| PEs | formulation | parameter scale |
| fully learnable APE (Gehring et al.,2017) | PxERD | L×D |
| fixed sinusoidal APE (Vaswani et al.,2017) | P(x)=[..,sin(wix）,cos(wix）,..]T； Wi=(1/10000)2i/D | 0 |
| learnable sinusoidal APE | P(x)=[...,sin(wix),cos(ωx)...]T; | D |
| fully learnable RPE | WiER PxERD | L×D |
| (Shaw et al.,2018) fixed sinusoidal RPE | P(x)=[.,sin(wix）,cos(wix）,T； | 0 |
| (Wei et al.,2019) learnable sinusoidal RPE | Wi=(1/10000)2i/D P(x)=[...,sin(ωix),cos(wx）,...]； WiER | L |
",
+ "bbox": [
+ 282,
+ 106,
+ 715,
+ 253
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Note that sinusoidal PEs satisfy both Property 2 (translation invariance) because the inner product is only associated with its position difference $x - y$ , and Property 3 (symmetry), because the dot product itself is symmetric: $\\langle \\vec { x } , \\vec { y } \\rangle = \\langle \\vec { y } , \\vec { x } \\rangle$ . Note also that checking Property 1 is equivalent to checking monotonicity oits first order derivative $\\begin{array} { r } { \\psi ( m ) = \\sum _ { i = 1 } ^ { D / 2 } \\cos ( \\omega _ { i } m ) } \\end{array}$ $\\psi ( m )$ is monotone on intervals wherehange sign, and these intervals $\\begin{array} { r } { \\psi ^ { \\prime } ( m ) = \\sum _ { i = 1 } ^ { D / 2 } - \\omega _ { i } \\sin ( \\omega _ { i } m ) } \\end{array}$ \ndepend on the choice of $\\omega _ { i }$ . With fixed frequencies $\\omega _ { i } = ( 1 / 1 0 0 0 0 ) ^ { 2 i / D }$ , it is monotonous when $m$ is roughly between 0 and 50, indicating that it can only strictly perceive a maximum distance of 50 and it is insensitive to faraway distances (e.g. longer than 50). ",
+ "bbox": [
+ 173,
+ 281,
+ 825,
+ 404
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Although sinusoidal PEs with fixed frequencies (i.e., $\\omega _ { i } = ( 1 / 1 0 0 0 0 ) ^ { 2 i / D } )$ are common in APEs and RPEs, we argue that learning these frequencies is useful because it can adaptively adjust intervals of monotonicity (they do not have to be 0-50 as in the fixed sinusoidal APE) 4. With trainable frequencies, we can adaptively allocate a number of frequencies in a data-driven way. App. A.2 explains the expressive power of sinusoidal PEs with trainable frequencies from the perspective of the Fourier series. Extending existing fixed sinusoidal PEs to a learnable version with learnable frequencies gives two variants: a learnable sinusoidal APE and a learnable sinusoidal RPE. ",
+ "bbox": [
+ 174,
+ 410,
+ 825,
+ 508
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "3.2 UNDERSTANDING RPES ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 526,
+ 382,
+ 541
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "RPEs ignore the absolute position of words and directly encode their relative distance. The RPEs expression adheres to the translation invariance property during parameterization, since relative distance with the same offset will be embedded as the same embedding, namely, $P _ { x _ { 1 } - y _ { 1 } } = P _ { x _ { 2 } - y _ { 2 } }$ if $x _ { 1 } - y _ { 1 } = x _ { 2 } - y _ { 2 }$ . Plus, RPEs that separately embed forward and backward relative embeddings, i.e., $P _ { i - j } \\neq P _ { j - i }$ , do not meet symmetry during parameterization. ",
+ "bbox": [
+ 173,
+ 553,
+ 825,
+ 625
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Sinusoidal RPEs can also embed neighboring relative position in close vectors with a local monotonicity, similarly to sinusoidal APEs. Note that the dot products between two sinusoidal relative position vectors with the same offset, without distinguishing positive negative relative position vectors, should be identical 5. This makes it hardly perceive of the border between preceding and succeeding relative position vectors. ",
+ "bbox": [
+ 174,
+ 630,
+ 825,
+ 700
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "4 EXAMINING PE PROPERTIES IN PRE-TRAINED LANGUAGE MODEL",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 723,
+ 746,
+ 738
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "We train BERT with six basic PEs as in Tab. 1 and their combination variants, and conduct a probing test to check to which degree they satisfy the properties. ",
+ "bbox": [
+ 171,
+ 755,
+ 823,
+ 784
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Pre-training The pre-trained “BERT-base-uncased” checkpoint (Devlin et al., 2018) is used to train by replacing the original absolute PE module with a new PE variant (including APEs and RPEs). We train the new models with a sequence length of 128 for 5 epochs and then 512 for another 2 epochs. The training is the same as in the original BERT, i.e., BooksCorpus and Wikipedia (16G raw documents) with whole word masking. To be fair, the BERT with the original fully-learnable ",
+ "bbox": [
+ 173,
+ 800,
+ 825,
+ 871
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "image",
+ "img_path": "images/30f1c450bd4b79e222786aa554bae24eb1aa15ffc26cb223cfab23b47e4800e1.jpg",
+ "image_caption": [
+ "Figure 1: Dot products between absolute position vectors 6(top row) and relative position vectors (bottom row). Darker means the two position vectors are closer. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 174,
+ 79,
+ 821,
+ 382
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "APE is also further trained in the same way. All models have about 110M parameters corresponding to a typical base setting, with minor differences solely depending on the parameterization in Tab. 1. ",
+ "bbox": [
+ 173,
+ 429,
+ 823,
+ 458
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4.1 DOT PRODUCT BETWEEN POSITION VECTORS ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 476,
+ 527,
+ 489
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "APEs We calculate dot products between two arbitrary position vectors for APEs and RPEs (see Fig. 1). For APEs, neighboring position vectors are generally closer compared to faraway ones. This trend is clearer in the learnable sinusoidal APE, which imposes a strict sinusoidal regularization for PEs. Note that additionally adopting RPEs does not affect too much PE patterns, as can be seen by comparing Fig. 1(a) and 1(b), or Fig. 1(c) and 1(d). ",
+ "bbox": [
+ 173,
+ 501,
+ 825,
+ 570
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "RPEs In the fully-learnable RPE setting, the vertical and horizontal bright bands in 1(e) and 1(f) show that the relative position vectors for small offsets (e.g., $\\{ P _ { - 5 } , \\cdot \\cdot \\cdot , P _ { 0 } , \\cdot \\cdot \\cdot P _ { 5 } \\}$ ) are notably different to other relative position vectors; it indicates that the relative position vectors with small offsets are more distinguishable than faraway relative position vectors. The four dark corners in 1(e) and 1(f) means that relative position vectors with longer offset than 20, i.e., from -64 to -20 and from 20 to 64, are very close, showing that the fully-learnable RPE does not significantly distinguish far-distant RPEs. This suggests that truncating RPEs into a fixed distance (e.g. 64 in (Shaw et al., 2018)), is reasonable. This effect is further explained in App. D. ",
+ "bbox": [
+ 173,
+ 585,
+ 825,
+ 698
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4.2 IDENTICAL WORD PROBING ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 715,
+ 405,
+ 728
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "In APEs, the attention matrix $( A = \\operatorname { s o f t m a x } ( Q K ^ { T } ) )$ is related to individual words and their positions, an element of (inactivated) $A$ in the first layer is given by: ",
+ "bbox": [
+ 169,
+ 739,
+ 821,
+ 770
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "img_path": "images/72f8b51657978e667d0519f2173fd7d615065093efddfb1c98cdad8d0e5a2583.jpg",
+ "text": "$$\n\\begin{array} { r l } & { a _ { i j } = ( w _ { i } + p _ { i } ) W ^ { Q , 1 } ( ( w _ { j } + p _ { j } ) W ^ { K , 1 } ) ^ { T } } \\\\ & { \\quad = \\underbrace { w _ { i } W ^ { Q , 1 } ( W ^ { K , 1 } ) ^ { T } w _ { j } ^ { T } } _ { \\mathrm { w o r d - w o r d ~ c o r e s p o n d e n c e } } + \\underbrace { w _ { i } W ^ { Q , 1 } ( W ^ { K , 1 } ) ^ { T } p _ { j } ^ { T } } _ { \\mathrm { w o r d - p o i n i m ~ c o r r e s p o n d e n c e } } + \\underbrace { p _ { i } W ^ { Q , 1 } ( W ^ { K , 1 } ) ^ { T } w _ { j } ^ { T } } _ { \\mathrm { w o r d - p o i n i m ~ c o r r e s p o n d e n c e } } + \\underbrace { p _ { i } W ^ { Q , 1 } ( W ^ { K , 1 } ) ^ { T } p _ { j } ^ { T } } _ { \\mathrm { p o s i t i o n - p o s i t i o n ~ c o r r e s p o n d e n c e } } } \\end{array}\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 184,
+ 773,
+ 790,
+ 833
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Identical word probing for PEs To study the effect of only PEs in $A$ without considering individual words, we use identical word probing: feed many repeated identical words (can be arbitrary, ",
+ "bbox": [
+ 173,
+ 845,
+ 821,
+ 875
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "image",
+ "img_path": "images/747fdde3ec02dea5b0654be29c91d4531195ea03942c0651b42adbabea11b8b6.jpg",
+ "image_caption": [
+ "Figure 2: Identical word probing. Darker in the $i$ -th row and $j$ -th column means that the $i$ -th words generally attend more on the $j$ -th words. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 176,
+ 78,
+ 821,
+ 369
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "denoted as $\\bar { w }$ ) as a sentence to BERT to check the attention values $\\bar { A } ^ { ( 1 ) }$ , with each element ",
+ "bbox": [
+ 171,
+ 405,
+ 769,
+ 420
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "equation",
+ "img_path": "images/e9995aba1be098039ea7e5a220b11bc27987167da58202a256ab11da5e1ac565.jpg",
+ "text": "$$\n\\bar { a } _ { i j } ^ { 1 } ( \\bar { w } ) = ( \\bar { w } + p _ { i } ) { W } ^ { Q , 1 } ( ( \\bar { w } + p _ { j } ) { W } ^ { K , 1 } ) ^ { T }\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 364,
+ 421,
+ 633,
+ 440
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "As we take an average of $\\bar { A } ^ { ( 1 ) }$ over many randomly-selected words $\\bar { w }$ , the general patterns of $\\bar { A } ^ { ( 1 ) }$ will not be affected by any particular word. Namely, $\\bar { A } ^ { ( 1 ) }$ is word-free and only related to learned PEs. Thus, $\\bar { A } ^ { ( 1 ) }$ can be treated as a general attention bias and can also implicitly convey positionwise proximity in Transformers. Note that the probing test could also be applied to RPEs. ",
+ "bbox": [
+ 174,
+ 450,
+ 825,
+ 511
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "4.2.1 QUALITATIVE ANALYSIS ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 525,
+ 397,
+ 539
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Fig. 2 shows the average attention weights among all heads in the first layer. BERT without $P E$ nearly treats all words uniformly (bag-of-words). Almost all APEs and RPEs have a clear pattern of translation invariance, local monotonicity in a neighboring window, and symmetry. Note that this is nontrivial since no specific constraints or priors were imposed on fully-learnable APEs/RPEs 7. ",
+ "bbox": [
+ 173,
+ 549,
+ 825,
+ 604
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "BERT with APEs does not show any direction awareness since Fig. 2(b) and 2(c) are nearly symmetrical. As seen from Fig. 2(f,h), BERT with learnable sinusoidal $R P E$ generally attends more on forward tokens than backward tokens, which cannot be clearly found in fully-learnable RPE and fixed sinusoidal RPE. Interestingly, the white bands along the diagonal in Fig. 2 (d, f, g) suggest that some words generally do not attend to themselves, as previously observed in (Clark et al., 2019) 8 . ",
+ "bbox": [
+ 173,
+ 611,
+ 825,
+ 683
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "4.2.2 QUANTITATIVE ANALYSIS ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 696,
+ 406,
+ 710
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Using the activated attention values $\\bar { A } ^ { ( 1 ) }$ in Eq. $8 ^ { 9 }$ we adopt three quantitative indicators to measure to which extent BERT models with individual PEs satisfy the three properties and their derivative indicators (see App. B for details of calculating these indicators) in Tab. 2. Basically, all APEs and RPEs satisfy monotonicity in small offsets and translation invariance compared to BERT without $P E$ ; All PEs nearly satisfy symmetry except for the learnable sinusoidal RPE and its combinations. ",
+ "bbox": [
+ 174,
+ 719,
+ 825,
+ 790
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "APEs and RPEs The learnable sinusoidal APE better satisfies all three properties than fully learnable APE and fixed sinusoidal APE; this is due to its sinusoidal parameterization and flexible frequencies. RPEs satisfy translation invariance to a higher degree than APEs, because they directly ",
+ "bbox": [
+ 176,
+ 804,
+ 821,
+ 847
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Table 2: Quantitative measurement of the properties (monotonicity, translation invariance, symmetry, and direction balance 10). For these property indicators, the smaller the number, the better the property is met. 0 denotes that the property is ideally satisfied. Direction balance denotes the ratio between the sum of attention values for forward attending and backward attending. 1 means it is fully-balanced in directions. We have indicated the numbers that most closely correspond to satisfaction properties and direction balance for each group in bold. ",
+ "bbox": [
+ 174,
+ 75,
+ 823,
+ 160
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "table",
+ "img_path": "images/9811d60faab6bac1a300b5f26278564c1facba4a79c7d13f19aa46f9e0905c55.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": " | Stages | 1 Stage | Notes |
| WaveNet (van den Oord et al.,2016) | AAu Ling | × | |
| SampleRNN (Mehri et al., 2017) | AAu | × | not a TTS model |
| Deep Voice (Arik et al.,2017) | Ch APhLing ARAu | × | uses segmentation model |
| WaveRNN (Kalchbrenner etal.,2018) | Ling AR ARAu | × | |
| LPCNet (Valin & Skoglund,2019) | ARAu Cep | × | |
| WaveGlow (Prenger et al.,2019) | MelS FF Au | × | |
| FloWaveNet (Kim et al.,2019) | FAu MelS | × | |
| WaveFlow (Ping et al.,2019b) | ARAu MelS | × | partially autoregressive |
| Par: WaveNet (van den Oord et al.,2018) | Ling FF* →Au | × | distillation |
| ClariNet (Ping et al.,2019a), teacher | ARAu Ch/Ph | √ | |
| ClariNet (Ping etal.,2019a),student | Au Ch/Ph | × | distillation |
| WaveGAN(Donahue et al.,2019) | Au | × | not a TTS model |
| MelGAN (Kumar etal., 2019) | MelsAu | × | |
| Par: WaveGAN(Yamamoto et al.,2020) | Ph AMels FAu | × | |
| AdVoc (Neekhara et al.,2019) | MelsMagS | × | |
| GAN-TTS (Binkowski etal., 2020) | Ling FAu FF | × | |
| Tacotron (Wang et al.,2017) | ChAMels MagS→Au | × | uses Griffin & Lim (1984) |
| Tacotron 2 (Shen et al.,2018) | AMelSAAu Ch | × | |
| Deep Voice 2(Gibiansky et al.,2017) | | × | uses segmentation model |
| DV2 Tacotron (Gibiansky et al.,2017) | ChAMagS AAu | × | |
| Deep Voice 3 (Ping et al.,2018) | Ch AMelS AAu | × | several alternative vocoders |
| TransformerTTS(Li etal.,2019) | Ch→Ph AMelS AAu | × | |
| Flowtron (Valle et al.,2020) | Ch AMelsAu | × | |
| VoiceLoop (Taigman etal.,2017) | Ph ALing → Au | × | |
| GAN Exposure (Guo et al.,2019) | Ph A MelS A Au | × | |
| MelNet (Vasquez & Lewis,2019) | AMelS→Au Ch- | × | |
| ParaNet (Peng et al.,2019) | FF* MelsAu Ch/Ph | × | distillation |
| FastSpeech (Ren et al.,2019) | FAu Ph FF* MelS | × | distillation |
| Flow-TTS (Miao et al.,2020) | Mels Au Ch | | |
| Glow-TTS(Kim et al.,2020) | PhMels Au | × × | |
| Char2wav (Sotelo et al.,2017) | ChALing ARAu | × | end-to-end finetuning |
| EATS (Ours) | Ch/Ph → Au | √ | |
+
+# J COMPARISON OF TTS METHODS
+
+In Table 6 we compare recent TTS approaches in terms of the inputs and outputs to each stage of the pipeline, and whether they are learnt in a single stage or multiple stages. Differentiating EATS from each prior approach is the fact that it learns a feed-forward mapping from text/phonemes to audio end-to-end in a single stage, without requiring distillation or separate vocoder training. The ClariNet teacher model (Ping et al., 2019a) is also trained in a single stage, but it uses teacher forcing to achieve this, requiring the model to be autoregressive. A separate distillation stage is necessary to obtain a feed-forward model in this case.
\ No newline at end of file
diff --git a/parse/train/rsf1z-JSj87/rsf1z-JSj87_content_list.json b/parse/train/rsf1z-JSj87/rsf1z-JSj87_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..1b9e2f4b66a7ca5ff99105f1eceb031d4f748b5b
--- /dev/null
+++ b/parse/train/rsf1z-JSj87/rsf1z-JSj87_content_list.json
@@ -0,0 +1,2611 @@
+[
+ {
+ "type": "text",
+ "text": "END-TO-END ADVERSARIAL TEXT-TO-SPEECH ",
+ "text_level": 1,
+ "bbox": [
+ 173,
+ 98,
+ 743,
+ 121
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Jeff Donahue∗, Sander Dieleman∗, Mikołaj Binkowski, Erich Elsen, Karen Simonyan ´ ∗ DeepMind {jeffdonahue,sedielem,binek,eriche,simonyan}@google.com ",
+ "bbox": [
+ 184,
+ 141,
+ 782,
+ 184
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "ABSTRACT ",
+ "text_level": 1,
+ "bbox": [
+ 454,
+ 220,
+ 544,
+ 236
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Modern text-to-speech synthesis pipelines typically involve multiple processing stages, each of which is designed or learnt independently from the rest. In this work, we take on the challenging task of learning to synthesise speech from normalised text or phonemes in an end-to-end manner, resulting in models which operate directly on character or phoneme input sequences and produce raw speech audio outputs. Our proposed generator is feed-forward and thus efficient for both training and inference, using a differentiable alignment scheme based on token length prediction. It learns to produce high fidelity audio through a combination of adversarial feedback and prediction losses constraining the generated audio to roughly match the ground truth in terms of its total duration and mel-spectrogram. To allow the model to capture temporal variation in the generated audio, we employ soft dynamic time warping in the spectrogram-based prediction loss. The resulting model achieves a mean opinion score exceeding 4 on a 5 point scale, which is comparable to the state-of-the-art models relying on multi-stage training and additional supervision.1 ",
+ "bbox": [
+ 232,
+ 252,
+ 764,
+ 460
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "1 INTRODUCTION ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 489,
+ 336,
+ 506
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "A text-to-speech (TTS) system processes natural language text inputs to produce synthetic human-like speech outputs. Typical TTS pipelines consist of a number of stages trained or designed independently – e.g. text normalisation, aligned linguistic featurisation, mel-spectrogram synthesis, and raw audio waveform synthesis (Taylor, 2009). Although these pipelines have proven capable of realistic and high-fidelity speech synthesis and enjoy wide real-world use today, these modular approaches come with a number of drawbacks. They often require supervision at each stage, in some cases necessitating expensive “ground truth” annotations to guide the outputs of each stage, and sequential training of the stages. Further, they are unable to reap the full potential rewards of data-driven “end-to-end\" learning widely observed in a number of prediction and synthesis task domains across machine learning. ",
+ "bbox": [
+ 173,
+ 521,
+ 825,
+ 646
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "In this work, we aim to simplify the TTS pipeline and take on the challenging task of synthesising speech from text or phonemes in an end-to-end manner. We propose EATS – End-to-end Adversarial Text-to-Speech – generative models for TTS trained adversarially (Goodfellow et al., 2014) that operate on either pure text or raw (temporally unaligned) phoneme input sequences, and produce raw speech waveforms as output. These models eliminate the typical intermediate bottlenecks present in most state-of-the-art TTS engines by maintaining learnt intermediate feature representations throughout the network. ",
+ "bbox": [
+ 174,
+ 650,
+ 825,
+ 748
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Our speech synthesis models are composed of two high-level submodules, detailed in Section 2. An aligner processes the raw input sequence and produces relatively low-frequency $( 2 0 0 \\ : \\mathrm { H z } )$ aligned features in its own learnt, abstract feature space. The features output by the aligner may be thought of as taking the place of the earlier stages of typical TTS pipelines – e.g., temporally aligned melspectrograms or linguistic features. These features are then input to the decoder which upsamples the features from the aligner by 1D convolutions to produce $2 4 \\mathrm { k H z }$ audio waveforms. ",
+ "bbox": [
+ 174,
+ 752,
+ 825,
+ 837
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "By carefully designing the aligner and guiding training by a combination of adversarial feedback and domain-specific loss functions, we demonstrate that a TTS system can be learnt nearly end-to-end, resulting in high-fidelity natural-sounding speech approaching the state-of-the-art TTS systems. Our main contributions include: ",
+ "bbox": [
+ 176,
+ 840,
+ 823,
+ 869
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "",
+ "bbox": [
+ 171,
+ 103,
+ 825,
+ 132
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "• A fully differentiable and efficient feed-forward aligner architecture that predicts the duration of each input token and produces an audio-aligned representation. \n• The use of flexible dynamic time warping-based prediction losses to enforce alignment with input conditioning while allowing the model to capture the variability of timing in human speech. \n• An overall system achieving a mean opinion score of 4.083, approaching the state of the art from models trained using richer supervisory signals. ",
+ "bbox": [
+ 173,
+ 141,
+ 826,
+ 238
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2 METHOD ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 256,
+ 282,
+ 272
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Our goal is to learn a neural network (the generator) which maps an input sequence of characters or phonemes to raw audio at $2 4 \\mathrm { k H z }$ . Beyond the vastly different lengths of the input and output signals, this task is also challenging because the input and output are not aligned, i.e. it is not known beforehand which output tokens each input token will correspond to. To address these challenges, we divide the generator into two blocks: (i) the aligner, which maps the unaligned input sequence to a representation which is aligned with the output, but has a lower sample rate of $2 0 0 \\mathrm { H z }$ ; and (ii) the decoder, which upsamples the aligner’s output to the full audio frequency. The entire generator architecture is differentiable, and is trained end to end. Importantly, it is also a feed-forward convolutional network, which makes it well-suited for applications where fast batched inference is important: our EATS implementation generates speech at a speed of $2 0 0 \\times$ realtime on a single NVIDIA V100 GPU (see Appendix A and Table 3 for details). It is illustrated in Figure 1. ",
+ "bbox": [
+ 173,
+ 285,
+ 825,
+ 439
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "The generator is inspired by GAN-TTS (Binkowski et al., 2020), a text-to-speech generative ad- ´ versarial network operating on aligned linguistic features. We employ the GAN-TTS generator as the decoder in our model, but instead of upsampling pre-computed linguistic features, its input comes from the aligner block. We make it speaker-conditional by feeding in a speaker embedding s alongside the latent vector $\\mathbf { z }$ , to enable training on a larger dataset with recordings from multiple speakers. We also adopt the multiple random window discriminators (RWDs) from GAN-TTS, which have been proven effective for adversarial raw waveform modelling, and we preprocess real audio input by applying a simple $\\mu$ -law transform. Hence, the generator is trained to produce audio in the $\\mu$ -law domain and we apply the inverse transformation to its outputs when sampling. ",
+ "bbox": [
+ 173,
+ 443,
+ 825,
+ 569
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "The loss function we use to train the generator is as follows: ",
+ "bbox": [
+ 173,
+ 573,
+ 568,
+ 588
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "equation",
+ "img_path": "images/27eac29c9a6d3953f508878899213ec3765f3d0e5217338928d3458581c3a01f.jpg",
+ "text": "$$\n\\mathcal { L } _ { G } = \\mathcal { L } _ { G , \\mathrm { a d v } } + \\lambda _ { \\mathrm { p r e d } } \\cdot \\mathcal { L } _ { \\mathrm { p r e d } } ^ { \\prime \\prime } + \\lambda _ { \\mathrm { l e n g t h } } \\cdot \\mathcal { L } _ { \\mathrm { l e n g t h } } ,\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 333,
+ 594,
+ 661,
+ 614
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "where $\\mathcal { L } _ { G , \\mathrm { a d v } }$ is the adversarial loss, linear in the discriminators’ outputs, paired with the hinge loss (Lim & Ye, 2017; Tran et al., 2017) used as the discriminators’ objective, as used in GANTTS (Binkowski et al., 2020). The use of an adversarial (Goodfellow et al., 2014) loss is an advantage ´ of our approach, as this setup allows for efficient feed-forward training and inference, and such losses tend to be mode-seeking in practice, a useful behaviour in a strongly conditioned setting where realism is an important design goal, as in the case of text-to-speech. In the remainder of this section, we describe the aligner network and the auxiliary predictiondetail, and recap the components which were adopted from G $( \\mathcal { L } _ { \\mathrm { { p r e d } } } ^ { \\prime \\prime } )$ and lengthS. $( \\mathcal { L } _ { \\mathrm { l e n g t h } } )$ losses in ",
+ "bbox": [
+ 173,
+ 619,
+ 825,
+ 731
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2.1 ALIGNER ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 747,
+ 281,
+ 761
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Given a token sequence ${ \\bf x } = ( x _ { 1 } , \\dots , x _ { N } )$ of length $N$ , we first compute token representations ${ \\bf h } = f ( { \\bf x } , { \\bf z } , { \\bf s } )$ , where $f$ is a stack of dilated convolutions (van den Oord et al., 2016) interspersed with batch normalisation (Ioffe & Szegedy, 2015) and ReLU activations. The latents $\\mathbf { z }$ and speaker embedding s modulate the scale and shift parameters of the batch normalisation layers (Dumoulin et al., 2017; De Vries et al., 2017). We then predict the length for each input token individually: $l _ { n } = g ( h _ { n } , \\mathbf { z } , \\mathbf { s } )$ , where $g$ is an MLP. We use a ReLU nonlinearity at the output to ensure that the predicted lengths are non-sum of the token lengths: $\\begin{array} { r } { e _ { n } = \\sum _ { m = 1 } ^ { n } l _ { m } } \\end{array}$ then find the predicted token end p, and the token centre positions as $\\begin{array} { r } { c _ { n } = e _ { n } - \\frac { 1 } { 2 } l _ { n } } \\end{array}$ mulative. Based on these predicted positions, we can interpolate the token representations into an audio-aligned representation at $2 0 0 \\mathrm { H z }$ , $\\mathbf { a } = ( a _ { 1 } , \\ldots , a _ { S } )$ , where $S = \\lceil e _ { N } \\rceil$ is the total number of output time steps. To compute $a _ { t }$ , we obtain interpolation weights for the token representations $h _ { n }$ using a softmax over the squared distance between $t$ and $c _ { n }$ , scaled by a temperature parameter $\\sigma ^ { 2 }$ , which we set to 10.0 (i.e. a Gaussian kernel): ",
+ "bbox": [
+ 173,
+ 770,
+ 826,
+ 924
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "",
+ "bbox": [
+ 173,
+ 103,
+ 825,
+ 132
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "equation",
+ "img_path": "images/45815cf2671e815e34f001fd64acc004229a13ba08900675bb14c40609d9b470.jpg",
+ "text": "$$\nw _ { t } ^ { n } = \\frac { \\exp { \\left( - \\sigma ^ { - 2 } ( t - c _ { n } ) ^ { 2 } \\right) } } { \\sum _ { m = 1 } ^ { N } \\exp { \\left( - \\sigma ^ { - 2 } ( t - c _ { m } ) ^ { 2 } \\right) } } .\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 372,
+ 133,
+ 625,
+ 174
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Using these weights, we can then compute $\\begin{array} { r } { a _ { t } \\ = \\ \\sum _ { n = 1 } ^ { N } w _ { t } ^ { n } h _ { n } } \\end{array}$ , which amounts to non-uniform interpolation. By predicting token lengths and obtaining positions using cumulative summation, instead of predicting positions directly, we implicitly enforce monotonicity of the alignment. Note that tokens which have a non-monotonic effect on prosody, such as punctuation, can still affect the entire utterance thanks to the stack of dilated convolutions $f$ , whose receptive field is large enough to allow for propagation of information across the entire token sequence. The convolutions also ensure generalisation across different sequence lengths. Appendix B includes pseudocode for the aligner. ",
+ "bbox": [
+ 173,
+ 180,
+ 825,
+ 281
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "2.2 WINDOWED GENERATOR TRAINING ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 299,
+ 460,
+ 313
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Training examples vary widely in length, from about 1 to 20 seconds. We cannot pad all sequences to a maximal length during training, as this would be wasteful and prohibitively expensive: 20 seconds of audio at $2 4 \\mathrm { k H z }$ correspond to 480,000 timesteps, which results in high memory requirements. Instead, we randomly extract a 2 second window from each example, which we will refer to as a training window, by uniformly sampling a random offset $\\eta$ . The aligner produces a $2 0 0 \\mathrm { H z }$ audio-aligned representation for this window, which is then fed to the decoder (see Figure 1). Note that we only need to compute $a _ { t }$ for time steps $t$ that fall within the sampled window, but we do have to compute the predicted token lengths $l _ { n }$ for the entire input sequence. During evaluation, we simply produce the audio-aligned representation for the full utterance and run the decoder on it, which is possible because it is fully convolutional. ",
+ "bbox": [
+ 174,
+ 324,
+ 485,
+ 601
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "2.3 ADVERSARIAL DISCRIMINATORS ",
+ "text_level": 1,
+ "bbox": [
+ 178,
+ 619,
+ 441,
+ 632
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "image",
+ "img_path": "images/e8346596cd53edabed8d93e517d344c6bdce1b87d2b7209f1d540d66775d362c.jpg",
+ "image_caption": [
+ "Figure 1: A diagram of the generator, including the monotonic interpolation-based aligner. $z$ and ch denote the latent Gaussian vector and the number of output channels, respectively. During training, audio windows have a fixed length of 2 seconds and are generated from the conditioning text using random offsets $\\eta$ and predicted phoneme lengths; the shaded areas in the logits grid and waveform are not synthesised. For inference (sampling), we set $\\eta = 0$ . In the No Phonemes ablation, the phonemizer is skipped and the character sequence is fed directly into the aligner. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 500,
+ 292,
+ 823,
+ 662
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Random window discriminators. We use an ensemble of random window discriminators (RWDs) adopted from GAN-TTS. Each RWD operates on audio fragments of different lengths, randomly sampled from the training window. We use five RWDs with window sizes 240, 480, 960, 1920 and 3600. This enables each RWD to operate at a different resolution. Note that 3600 samples at $2 4 ~ \\mathrm { k H z }$ corresponds to $1 5 0 ~ \\mathrm { m s }$ of audio, so all RWDs operate on short timescales. All RWDs in our model are unconditional with respect to text: they cannot access the text sequence or the aligner output. (GAN-TTS uses 10 RWDs, including 5 conditioned on linguistic features which we omit.) They are, however, conditioned on the speaker, via projection embedding (Miyato & Koyama, 2018). ",
+ "bbox": [
+ 174,
+ 643,
+ 485,
+ 851
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "",
+ "bbox": [
+ 176,
+ 852,
+ 705,
+ 866
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Spectrogram discriminator. We use an additional discriminator which operates on the full training window in the spectrogram domain. We extract log-scaled mel-spectrograms from the audio signals and use the BigGAN-deep architecture (Brock et al., 2018), essentially treating the spectrograms as ",
+ "bbox": [
+ 176,
+ 882,
+ 823,
+ 924
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "images. The spectrogram discriminator also uses speaker identity through projection embedding. \nDetails on the spectrogram discriminator architecture are included in Appendix C. ",
+ "bbox": [
+ 171,
+ 103,
+ 825,
+ 132
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "2.4 SPECTROGRAM PREDICTION LOSS ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 147,
+ 449,
+ 161
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "In preliminary experiments, we discovered that adversarial feedback is insufficient to learn alignment. At the start of training, the aligner does not produce an accurate alignment, so the information in the input tokens is incorrectly temporally distributed. This encourages the decoder to ignore the aligner output. The unconditional discriminators provide no useful learning signal to correct this. If we want to use conditional discriminators instead, we face a different problem: we do not have aligned ground truth. Conditional discriminators also need an aligner module, which cannot function correctly at the start of training, effectively turning them into unconditional discriminators. Although it should be possible in theory to train the discriminators’ aligner modules adversarially, we find that this does not work in practice, and training gets stuck. ",
+ "bbox": [
+ 173,
+ 170,
+ 825,
+ 295
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Instead, we propose to guide learning by using an explicit prediction loss in the spectrogram domain: we minimise the $L _ { 1 }$ loss between the log-scaled mel-spectrograms of the generator output, and the corresponding ground truth training window. This helps training to take off, and renders conditional discriminators unnecessary, simplifying the model. Let $S _ { \\mathrm { g e n } }$ be the spectrogram of the generated audio, $S _ { \\mathrm { g t } }$ the spectrogram of the corresponding ground truth, and $S [ t , f ]$ the log-scaled magnitude at time step $t$ and mel-frequency bin $f$ . Then the prediction loss is: ",
+ "bbox": [
+ 173,
+ 299,
+ 825,
+ 383
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "img_path": "images/dfef56075b01fa14853b720235db6f8454bc544cb2c4c329b0b4af381fd1c3ba.jpg",
+ "text": "$$\n\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { p r e d } } = \\frac { 1 } { F } \\sum _ { t = 1 } ^ { T } \\sum _ { f = 1 } ^ { F } | S _ { \\mathrm { g e n } } [ t , f ] - S _ { \\mathrm { g t } } [ t , f ] | . } \\end{array}\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 341,
+ 397,
+ 658,
+ 420
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "$T$ and $F$ are the total number of time steps and mel-frequency bins respectively. Computing the prediction loss in the spectrogram domain, rather than the time domain, has the advantage of increased invariance to phase differences between the generated and ground truth signals, which are not perceptually salient. Seeing as the spectrogram extraction operation has several hyperparameters and its implementation is not standardised, we provide the code we used for this in Appendix D. We applied a small amount of jitter (by up to $\\pm 6 0$ samples at $2 4 \\mathrm { k H z }$ ) to the ground truth waveform before computing $S _ { \\mathrm { g t } }$ , which helped to reduce artifacts in the generated audio. ",
+ "bbox": [
+ 173,
+ 426,
+ 825,
+ 526
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "The inability to learn alignment from adversarial feedback alone is worth expanding on: likelihoodbased autoregressive models have no issues learning alignment, because they are able to benefit from teacher forcing (Williams & Zipser, 1989) during training: the model is trained to perform next step prediction on each sequence step, given the preceding ground truth, and it is expected to infer alignment only one step at a time. This is not compatible with feed-forward adversarial models however, so the prediction loss is necessary to bootstrap alignment learning for our model. ",
+ "bbox": [
+ 174,
+ 530,
+ 825,
+ 614
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Note that although we make use of mel-spectrograms for training in $\\mathcal { L } _ { \\mathrm { p r e d } }$ (and to compute the inputs for the spectrogram discriminator, Section 2.3), the generator itself does not produce spectrograms as part of the generation process. Rather, its outputs are raw waveforms, and we convert these waveforms to spectrograms only for training (backpropagating gradients through the waveform to mel-spectrogram conversion operation). ",
+ "bbox": [
+ 174,
+ 618,
+ 825,
+ 688
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "2.5 DYNAMIC TIME WARPING ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 703,
+ 392,
+ 717
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "The spectrogram prediction loss incorrectly assumes that token lengths are deterministic. We can relax the requirement that the generated and ground truth spectrograms are exactly aligned, by incorporating dynamic time warping (DTW) (Sakoe, 1971; Sakoe & Chiba, 1978). We calculate the prediction loss by iteratively finding a minimal-cost alignment path $p$ between the generated and target spectrograms, $S _ { \\mathrm { g e n } }$ and $S _ { \\mathrm { g t } }$ . We start at the first time step in both spectrograms: $p _ { \\mathrm { g e n , 1 } } = p _ { \\mathrm { g t , 1 } } = 1$ At each iteration $k$ , we take one of three possible actions: ",
+ "bbox": [
+ 174,
+ 726,
+ 825,
+ 810
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "1. go to the next time step in both $S _ { \\mathrm { g e n } } , S _ { \\mathrm { g t } } \\colon p _ { \\mathrm { g e n } , k + 1 } = p _ { \\mathrm { g e n } , k } + 1 , p _ { \\mathrm { g t } , k + 1 } = p _ { \\mathrm { g t } , k } + 1 ;$ \n2. go to the next time step in $S _ { \\mathrm { g t } }$ only: $p _ { \\mathrm { g e n } , k + 1 } = p _ { \\mathrm { g e n } , k } , p _ { \\mathrm { g t } , k + 1 } = p _ { \\mathrm { g t } , k } + 1$ ; \n3. go to the next time step in $S _ { \\mathrm { g e n } }$ only: $p _ { \\mathrm { g e n } , k + 1 } = p _ { \\mathrm { g e n } , k } + 1$ , pgt,k+1 = pgt,k. ",
+ "bbox": [
+ 212,
+ 819,
+ 805,
+ 875
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "The resulting path is $p = \\langle ( p _ { \\mathrm { g e n } , 1 } , p _ { \\mathrm { g t } , 1 } ) , \\dots , ( p _ { \\mathrm { g e n } , K _ { p } } , p _ { \\mathrm { g t } , K _ { p } } ) \\rangle$ , where $K _ { p }$ is the length. Each action is assigned a cost based on the $L _ { 1 }$ distance between $S _ { \\mathrm { g e n } } [ p _ { \\mathrm { g e n } , k } ]$ and $\\mathrm { \\dot { \\cal S } } _ { \\mathrm { g t } } [ p _ { \\mathrm { g t } , k } ]$ , and a warp penalty $w$ which is incurred if we choose not to advance both spectrograms in lockstep (i.e. we are warping the spectrogram by taking action 2 or 3; we use $w = 1 . 0$ ). The warp penalty thus encourages alignment paths that do not deviate too far from the identity alignment. Let $\\delta _ { k }$ be an indicator which is 1 for iterations where warping occurs, and 0 otherwise. Then the total path cost $c _ { p }$ is: ",
+ "bbox": [
+ 176,
+ 881,
+ 821,
+ 924
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "",
+ "bbox": [
+ 173,
+ 103,
+ 825,
+ 146
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "img_path": "images/5e55027ea6418ddbbb991219c835af82b42c74a3172fd85b36bb0223dc35e50a.jpg",
+ "text": "$$\n\\begin{array} { r } { c _ { p } = \\sum _ { k = 1 } ^ { K _ { p } } \\Big ( \\boldsymbol { w } \\cdot \\boldsymbol { \\delta _ { k } } + \\frac { 1 } { F } \\sum _ { f = 1 } ^ { F } | S _ { \\mathrm { g e n } } [ p _ { \\mathrm { g e n } , k } , f ] - S _ { \\mathrm { g t } } [ p _ { \\mathrm { g t } , k } , f ] | \\Big ) . } \\end{array}\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 281,
+ 148,
+ 717,
+ 176
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "depends on the degree of warping $( T \\leq K _ { p } \\leq 2 T - 1 )$ . The DTW prediction loss is then: ",
+ "bbox": [
+ 192,
+ 178,
+ 781,
+ 194
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "img_path": "images/a0274c1bfcce379ed130b205b04215c357f6e780ecd4665614c4bcf802543ff3.jpg",
+ "text": "$$\n\\mathcal { L } _ { \\mathrm { p r e d } } ^ { \\prime } = \\operatorname* { m i n } _ { p \\in \\mathcal { P } } c _ { p } ,\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 442,
+ 196,
+ 553,
+ 219
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where $\\mathcal { P }$ is the set of all valid paths. $p \\in \\mathcal P$ only when $p _ { \\mathrm { g e n , 1 } } = p _ { \\mathrm { g t , 1 } } = 1$ and $p _ { \\mathrm { g e n } , K _ { p } } = p _ { \\mathrm { g t } , K _ { p } } = T$ i.e. the first and last timesteps of the spectrograms are aligned. To find the minimum, we use dynamic programming. Figure 2 shows a diagram of an optimal alignment path between two sequences. ",
+ "bbox": [
+ 173,
+ 223,
+ 825,
+ 265
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "DTW is differentiable, but the minimum across all paths makes optimisation difficult, because the gradient is propagated only through the minimal path. We use a soft version of DTW instead (Cuturi & Blondel, 2017), which replaces the minimum with the soft minimum: ",
+ "bbox": [
+ 173,
+ 270,
+ 614,
+ 325
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "img_path": "images/63b0bbdd0ea571b873711903c5b179bb9f83e424aee2ca04f4637076877dcb55.jpg",
+ "text": "$$\n\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { p r e d } } ^ { \\prime \\prime } = - \\tau \\cdot \\log \\sum _ { p \\in \\mathcal { P } } \\exp \\left( - \\frac { c _ { p } } { \\tau } \\right) , } \\end{array}\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 271,
+ 328,
+ 514,
+ 348
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where $\\tau = 0 . 0 1$ is a temperature parameter and the loss scale factor $\\lambda _ { \\mathrm { p r e d } } = 1 . 0$ . Note that the minimum operation is recovered by letting $\\tau 0$ . The resulting loss is a weighted aggregated cost across all paths, enabling gradient propagation through all feasible paths. This creates a trade-off: a higher $\\tau$ makes optimisation easier, but the resulting loss less accurately reflects the minimal path cost. Pseudocode for the soft DTW procedure is provided in Appendix E. ",
+ "bbox": [
+ 174,
+ 349,
+ 614,
+ 448
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "By relaxing alignment in the prediction loss, the generator can produce waveforms that are not exactly aligned, without being heavily penalised for it. This creates a synergy with the adversarial loss: instead of working against each other because of the rigidity of the prediction loss, the losses now cooperate to reward realistic audio generation with stochastic alignment. Note that the prediction loss is computed on a training window, and not on full length utterances, so we still assume that the start and end points of the windows are exactly aligned. While this might be incorrect, it does not seem to be much of a problem in practice. ",
+ "bbox": [
+ 174,
+ 452,
+ 617,
+ 536
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "image",
+ "img_path": "images/3ec946c31ee94122d585bbdb325279673d19931189642e35e802f93e8666c4d4.jpg",
+ "image_caption": [
+ "Figure 2: Dynamic time warping between two sequences finds a minimal-cost alignment path. Positions where warping occurs are marked with a border. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 630,
+ 286,
+ 823,
+ 436
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "",
+ "bbox": [
+ 173,
+ 536,
+ 818,
+ 577
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "2.6 ALIGNER LENGTH LOSS ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 590,
+ 380,
+ 604
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "To ensure that the model produces realistic token length predictions, we add a loss which encourages the predicted utterance length to be close to the ground truth length. This length is found by summing all token length predictions. Let $L$ be the the number of time steps in the training utterance at $2 0 0 \\mathrm { H z }$ , $l _ { n }$ the predicted length of the $n$ th token, and $N$ the number of tokens, then the length loss is: ",
+ "bbox": [
+ 173,
+ 614,
+ 825,
+ 671
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "img_path": "images/cfe00b01bfe2a11b352c8e3753171589565351b9c39d23dd72a5cfe5844cb4d9.jpg",
+ "text": "$$\n\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { l e n g t h } } = \\frac { 1 } { 2 } \\left( L - \\sum _ { n = 1 } ^ { N } l _ { n } \\right) ^ { 2 } . } \\end{array}\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 393,
+ 672,
+ 604,
+ 703
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "We use a scale factor $\\lambda _ { \\mathrm { l e n g t h } } = 0 . 1$ . Note that we cannot match the predicted lengths $l _ { n }$ to the ground truth lengths individually, because the latter are not available. ",
+ "bbox": [
+ 174,
+ 705,
+ 825,
+ 733
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "2.7 TEXT PRE-PROCESSING ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 747,
+ 379,
+ 761
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Although our model works well with character input, we find that sample quality improves significantly using phoneme input instead. This is not too surprising, given the heterogeneous way in which spellings map to phonemes, particularly in the English language. Many character sequences also have special pronunciations, such as numbers, dates, units of measurement and website domains, and a very large training dataset would be required for the model to learn to pronounce these correctly. Text normalisation (Zhang et al., 2019) can be applied beforehand to spell out these sequences as they are typically pronounced (e.g., 1976 could become nineteen seventy six), potentially followed by conversion to phonemes. We use an open source tool, phonemizer (Bernard, 2020), which performs partial normalisation and phonemisation (see Appendix F). Finally, whether we train on text or phoneme input sequences, we pre- and post-pad the sequence with a special silence token (for training and inference), to allow the aligner to account for silence at the beginning and end of each utterance. ",
+ "bbox": [
+ 173,
+ 770,
+ 826,
+ 924
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3 RELATED WORK ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 102,
+ 339,
+ 117
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Speech generation saw significant quality improvements once treating it as a generative modelling problem became the norm (Zen et al., 2009; van den Oord et al., 2016). Likelihood-based approaches dominate, but generative adversarial networks (GANs) (Goodfellow et al., 2014) have been making significant inroads recently. A common thread through most of the literature is a separation of the speech generation process into multiple stages: coarse-grained temporally aligned intermediate representations, such as mel-spectrograms, are used to divide the task into more manageable subproblems. Many works focus exclusively on either spectrogram generation or vocoding (generating a waveform from a spectrogram). Our work is different in this respect, and we will point out which stages of the generation process are addressed by each model. In Appendix J, Table 6 we compare these methods in terms of the inputs and outputs to each stage of their pipelines. ",
+ "bbox": [
+ 174,
+ 133,
+ 825,
+ 272
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Initially, most likelihood-based models for TTS were autoregressive (van den Oord et al., 2016; Mehri et al., 2017; Arik et al., 2017), which means that there is a sequential dependency between subsequent time steps of the produced output signal. That makes these models impractical for real-time use, although this can be addressed with careful engineering (Kalchbrenner et al., 2018; Valin & Skoglund, 2019). More recently, flow-based models (Papamakarios et al., 2019) have been explored as a feed-forward alternative that enables fast inference (without sequential dependencies). These can either be trained directly using maximum likelihood (Prenger et al., 2019; Kim et al., 2019; Ping et al., 2019b), or through distillation from an autoregressive model (van den Oord et al., 2018; Ping et al., 2019a). All of these models produce waveforms conditioned on an intermediate representation: either spectrograms or “linguistic features”, which contain temporally-aligned high-level information about the speech signal. Spectrogram-conditioned waveform models are often referred to as vocoders. ",
+ "bbox": [
+ 174,
+ 276,
+ 826,
+ 429
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "A growing body of work has applied GAN (Goodfellow et al., 2014) variants to speech synthesis (Donahue et al., 2019). An important advantage of adversarial losses for TTS is a focus on realism over diversity; the latter is less important in this setting. This enables a more efficient use of capacity compared to models trained with maximum likelihood. MelGAN (Kumar et al., 2019) and Parallel WaveGAN (Yamamoto et al., 2020) are adversarial vocoders, producing raw waveforms from mel-spectrograms. Neekhara et al. (2019) predict magnitude spectrograms from mel-spectrograms. Most directly related to our work is GAN-TTS (Binkowski et al., 2020), which produces waveforms ´ conditioned on aligned linguistic features, and we build upon that work. ",
+ "bbox": [
+ 174,
+ 434,
+ 825,
+ 545
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Another important line of work covers spectrogram generation from text. Such models rely on a vocoder to convert the spectrograms into waveforms (for which one of the previously mentioned models could be used, or a traditional spectrogram inversion technique (Griffin & Lim, 1984)). Tacotron 1 & 2 (Wang et al., 2017; Shen et al., 2018), Deep Voice 2 & 3 (Gibiansky et al., 2017; Ping et al., 2018), TransformerTTS (Li et al., 2019), Flowtron (Valle et al., 2020), and VoiceLoop (Taigman et al., 2017) are autoregressive models that generate spectrograms or vocoder features frame by frame. Guo et al. (2019) suggest using an adversarial loss to reduce exposure bias (Bengio et al., 2015; Ranzato et al., 2016) in such models. MelNet (Vasquez & Lewis, 2019) is autoregressive over both time and frequency. ParaNet (Peng et al., 2019) and FastSpeech (Ren et al., 2019) are nonautoregressive, but they require distillation (Hinton et al., 2015) from an autoregressive model. Recent flow-based approaches Flow-TTS (Miao et al., 2020) and Glow-TTS (Kim et al., 2020) are feedforward without requiring distillation. Most spectrogram generation models require training of a custom vocoder model on generated spectrograms, because their predictions are imperfect and the vocoder needs to be able to compensate for this2. Note that some of these works also propose new vocoder architectures in tandem with spectrogram generation models. ",
+ "bbox": [
+ 174,
+ 550,
+ 825,
+ 758
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Unlike all of the aforementioned methods, as highlighted in Appendix J, Table 6, our model is a single feed-forward neural network, trained end-to-end in a single stage, which produces waveforms given character or phoneme sequences, and learns to align without additional supervision from auxiliary sources (e.g. temporally aligned linguistic features from an external model) or teacher forcing. This simplifies the training process considerably. Char2wav (Sotelo et al., 2017) is finetuned end-to-end in the same fashion, but requires a pre-training stage with vocoder features used for intermediate supervision. ",
+ "bbox": [
+ 174,
+ 762,
+ 825,
+ 859
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Spectrogram prediction losses have been used extensively for feed-forward audio prediction models (Yamamoto et al., 2019; 2020; Yang et al., 2020; Arık et al., 2018; Engel et al., 2020; Wang et al., ",
+ "bbox": [
+ 176,
+ 864,
+ 823,
+ 893
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "2019; Défossez et al., 2018). We note that the $L _ { 1 }$ loss we use (along with (Défossez et al., 2018)), is comparatively simple, as spectrogram losses in the literature tend to have separate terms penalising magnitudes, log-magnitudes and phase components, each with their own scaling factors, and often across multiple resolutions. Dynamic time warping on spectrograms is a component of many speech recognition systems (Sakoe, 1971; Sakoe & Chiba, 1978), and has also been used for evaluation of TTS systems (Sailor & Patil, 2014; Chevelu et al., 2015). Cuturi & Blondel (2017) proposed the soft version of DTW we use in this work as a differentiable loss function for time series models. Kim et al. (2020) propose Monotonic Alignment Search (MAS), which relates to DTW in that both use dynamic programming to implicitly align sequences for TTS. However, they have different goals: MAS finds the optimal alignment between the text and a latent representation, whereas we use DTW to relax the constraints imposed by our spectrogram prediction loss term. Several mechanisms have been proposed to exploit monotonicity in tasks that require sequence alignment, including attention mechanisms (Graves, 2013; Zhang et al., 2018; Vasquez & Lewis, 2019; He et al., 2019; Raffel et al., 2017; Chiu & Raffel, 2018), loss functions (Graves et al., 2006; Graves, 2012) and search-based approaches (Kim et al., 2020). For TTS, incorporating this constraint has been found to help generalisation to long sequences (Battenberg et al., 2020). We incorporate monotonicity by using an interpolation mechanism, which is cheap to compute because it is not recurrent (unlike many monotonic attention mechanisms). ",
+ "bbox": [
+ 174,
+ 104,
+ 825,
+ 353
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "4 EVALUATION ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 371,
+ 313,
+ 386
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "In this section we discuss the setup and results of our empirical evaluation, describing the hyperparameter settings used for training and validating the architectural decisions and loss function components detailed in Section 2. Our primary metric used to evaluate speech quality is the Mean Opinion Score (MOS) given by human raters, computed by taking the mean of 1-5 naturalness ratings given across 1000 held-out conditioning sequences. In Appendix I we also report the Fréchet DeepSpeech Distance (FDSD), proposed by Binkowski et al. (2020) as a speech synthesis quality metric. Appendix A ´ reports training and evaluation hyperparameters we used for all experiments. ",
+ "bbox": [
+ 174,
+ 398,
+ 825,
+ 496
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "4.1 MULTI-SPEAKER DATASET ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 511,
+ 397,
+ 523
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "We train all models on a private dataset that consists of high-quality recordings of human speech performed by professional voice actors, and corresponding text. The voice pool consists of 69 female and male voices of North American English speakers, while the audio clips contain full sentences of lengths varying from less than 1 to 20 seconds at $2 4 \\mathrm { k H z }$ frequency. Individual voices are unevenly distributed, accounting for from 15 minutes to over 51 hours of recorded speech, totalling 260.49 hours. At training time, we sample 2 second windows from the individual clips, post-padding those shorter than 2 seconds with silence. For evaluation, we focus on the single most prolific speaker in our dataset, with all our main MOS results reported with the model conditioned on that speaker ID, but also report MOS results for each of the top four speakers using our main multi-speaker model. ",
+ "bbox": [
+ 174,
+ 534,
+ 825,
+ 659
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "4.2 RESULTS ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 672,
+ 277,
+ 686
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "In Table 1 we present quantitative results for our EATS model described in Section 2, as well as several ablations of the different model and learning signal components. The architecture and training setup of each ablation is identical to our base EATS model except in terms of the differences described by the columns in Table 1. Each ablation is “subtractive”, representing the full EATS system minus one particular feature. Our main result achieved by the base multi-speaker model is a mean opinion score (MOS) of 4.083. Although it is difficult to compare directly with prior results from the literature due to dataset differences, we nonetheless include MOS results from prior works (Binkowski et al., ´ 2020; van den Oord et al., 2016; 2018), with MOS in the 4.2 to $4 . 4 +$ range. Compared to these prior models, which rely on aligned linguistic features, EATS uses substantially less supervision. ",
+ "bbox": [
+ 174,
+ 695,
+ 825,
+ 821
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "The No RWDs, No MelSpecD, and No Discriminators ablations all achieved substantially worse MOS results than our proposed model, demonstrating the importance of adversarial feedback. In particular, the No RWDs ablation, with an MOS of 2.526, demonstrates the importance of the raw audio feedback, and removing RWDs significantly degrades the high frequency components. No MelSpecD causes intermittent artifacts and distortion, and removing all discriminators results in audio that sounds robotic and distorted throughout. The No $\\mathcal { L } _ { \\mathrm { l e n g t h } }$ and No $\\mathcal { L } _ { \\mathrm { p r e d } }$ ablations result in a model that does not train at all. Comparing our model with No DTW (MOS 3.559), the temporal flexibility provided by dynamic time warping significantly improves fidelity: removing it causes warbling and unnatural phoneme lengths. No Phonemes is trained with raw character inputs and attains MOS 3.423, due to occasional mispronunciations and unusual stress patterns. No Mon. Int. uses an aligner with a transformer-based attention mechanism (described in Appendix G) in place of our monotonic interpolation architecture, which turns out to generalise poorly to long utterances (yielding MOS 3.551). Finally, comparing against training with only a Single Speaker (MOS 3.829) shows that our EATS model benefits from a much larger multi-speaker dataset, even though MOS is evaluated only on this same single speaker on which the ablation was solely trained. Samples from each ablation are available at https://deepmind.com/research/publications/ End-to-End-Adversarial-Text-to-Speech. ",
+ "bbox": [
+ 174,
+ 825,
+ 825,
+ 924
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "table",
+ "img_path": "images/8e6caf3ffb57b432925eff74463cd565e20e83507b07d7669c740ecf637b01bf.jpg",
+ "table_caption": [
+ "Table 1: Mean Opinion Scores (MOS) for our final EATS model and the ablations described in Section 4, sorted by MOS. The middle columns indicate which components of our final model are enabled or ablated. Data describes the training set as Multispeaker (MS) or Single Speaker (SS). Inputs describes the inputs as raw characters (Ch) or phonemes $\\mathrm { ( P h ) }$ produced by Phonemizer. RWD (Random Window Discriminators), $M S D$ (Mel-spectrogram Discriminator), and $\\mathcal { L } _ { \\mathrm { l e n g t h } }$ (length prediction loss) indicate the presence $( \\checkmark )$ or absence $( \\times )$ of each of these training components described in Section 2. $\\mathcal { L } _ { \\mathrm { p r e d } }$ indicates which spectrogram prediction loss was used: with DTW $( \\mathcal { L } _ { \\mathrm { p r e d } } ^ { \\prime \\prime }$ , Eq. 6), without DTW $\\scriptstyle \\sum _ { \\mathrm { p r e d } }$ , Eq. 3), or absent $( \\times )$ . Align describes the architecture of the aligner as monotonic interpolation (MI) or attention-based (Attn). We also compare against recent state-of-the-art approaches from the literature which are trained on aligned linguistic features (unlike our models). Our MOS evaluation set matches that of GAN-TTS (Binkowski et al., 2020) (and our ´ “Single Speaker” training subset matches the GAN-TTS training set); the other approaches are not directly comparable due to dataset differences. "
+ ],
+ "table_footnote": [],
+ "table_body": " | Stages | 1 Stage | Notes |
| WaveNet (van den Oord et al.,2016) | AAu Ling | × | |
| SampleRNN (Mehri et al., 2017) | AAu | × | not a TTS model |
| Deep Voice (Arik et al.,2017) | Ch APhLing ARAu | × | uses segmentation model |
| WaveRNN (Kalchbrenner etal.,2018) | Ling AR ARAu | × | |
| LPCNet (Valin & Skoglund,2019) | ARAu Cep | × | |
| WaveGlow (Prenger et al.,2019) | MelS FF Au | × | |
| FloWaveNet (Kim et al.,2019) | FAu MelS | × | |
| WaveFlow (Ping et al.,2019b) | ARAu MelS | × | partially autoregressive |
| Par: WaveNet (van den Oord et al.,2018) | Ling FF* →Au | × | distillation |
| ClariNet (Ping et al.,2019a), teacher | ARAu Ch/Ph | √ | |
| ClariNet (Ping etal.,2019a),student | Au Ch/Ph | × | distillation |
| WaveGAN(Donahue et al.,2019) | Au | × | not a TTS model |
| MelGAN (Kumar etal., 2019) | MelsAu | × | |
| Par: WaveGAN(Yamamoto et al.,2020) | Ph AMels FAu | × | |
| AdVoc (Neekhara et al.,2019) | MelsMagS | × | |
| GAN-TTS (Binkowski etal., 2020) | Ling FAu FF | × | |
| Tacotron (Wang et al.,2017) | ChAMels MagS→Au | × | uses Griffin & Lim (1984) |
| Tacotron 2 (Shen et al.,2018) | AMelSAAu Ch | × | |
| Deep Voice 2(Gibiansky et al.,2017) | | × | uses segmentation model |
| DV2 Tacotron (Gibiansky et al.,2017) | ChAMagS AAu | × | |
| Deep Voice 3 (Ping et al.,2018) | Ch AMelS AAu | × | several alternative vocoders |
| TransformerTTS(Li etal.,2019) | Ch→Ph AMelS AAu | × | |
| Flowtron (Valle et al.,2020) | Ch AMelsAu | × | |
| VoiceLoop (Taigman etal.,2017) | Ph ALing → Au | × | |
| GAN Exposure (Guo et al.,2019) | Ph A MelS A Au | × | |
| MelNet (Vasquez & Lewis,2019) | AMelS→Au Ch- | × | |
| ParaNet (Peng et al.,2019) | FF* MelsAu Ch/Ph | × | distillation |
| FastSpeech (Ren et al.,2019) | FAu Ph FF* MelS | × | distillation |
| Flow-TTS (Miao et al.,2020) | Mels Au Ch | | |
| Glow-TTS(Kim et al.,2020) | PhMels Au | × × | |
| Char2wav (Sotelo et al.,2017) | ChALing ARAu | × | end-to-end finetuning |
| EATS (Ours) | Ch/Ph → Au | √ | |
",
+ "bbox": [
+ 179,
+ 172,
+ 810,
+ 702
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "text",
+ "text": "J COMPARISON OF TTS METHODS ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 102,
+ 472,
+ 117
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "In Table 6 we compare recent TTS approaches in terms of the inputs and outputs to each stage of the pipeline, and whether they are learnt in a single stage or multiple stages. Differentiating EATS from each prior approach is the fact that it learns a feed-forward mapping from text/phonemes to audio end-to-end in a single stage, without requiring distillation or separate vocoder training. The ClariNet teacher model (Ping et al., 2019a) is also trained in a single stage, but it uses teacher forcing to achieve this, requiring the model to be autoregressive. A separate distillation stage is necessary to obtain a feed-forward model in this case. ",
+ "bbox": [
+ 173,
+ 131,
+ 825,
+ 228
+ ],
+ "page_idx": 22
+ }
+]
\ No newline at end of file
diff --git a/parse/train/rsf1z-JSj87/rsf1z-JSj87_middle.json b/parse/train/rsf1z-JSj87/rsf1z-JSj87_middle.json
new file mode 100644
index 0000000000000000000000000000000000000000..c8c637288157168b6699cc298ee2f9ec7a77f468
--- /dev/null
+++ b/parse/train/rsf1z-JSj87/rsf1z-JSj87_middle.json
@@ -0,0 +1,60666 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 78,
+ 455,
+ 96
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 456,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 456,
+ 98
+ ],
+ "score": 1.0,
+ "content": "END-TO-END ADVERSARIAL TEXT-TO-SPEECH",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 113,
+ 112,
+ 479,
+ 146
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 112,
+ 480,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 112,
+ 480,
+ 126
+ ],
+ "score": 1.0,
+ "content": "Jeff Donahue∗, Sander Dieleman∗, Mikołaj Binkowski, Erich Elsen, Karen Simonyan ´ ∗",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 111,
+ 124,
+ 158,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 124,
+ 158,
+ 136
+ ],
+ "score": 1.0,
+ "content": "DeepMind",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 134,
+ 444,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 134,
+ 444,
+ 149
+ ],
+ "score": 1.0,
+ "content": "{jeffdonahue,sedielem,binek,eriche,simonyan}@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 175,
+ 333,
+ 187
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 174,
+ 335,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 174,
+ 335,
+ 189
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 142,
+ 200,
+ 468,
+ 365
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 199,
+ 470,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 199,
+ 470,
+ 214
+ ],
+ "score": 1.0,
+ "content": "Modern text-to-speech synthesis pipelines typically involve multiple processing",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 212,
+ 470,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 212,
+ 470,
+ 223
+ ],
+ "score": 1.0,
+ "content": "stages, each of which is designed or learnt independently from the rest. In this",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 223,
+ 469,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 223,
+ 469,
+ 234
+ ],
+ "score": 1.0,
+ "content": "work, we take on the challenging task of learning to synthesise speech from",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 233,
+ 470,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 233,
+ 470,
+ 245
+ ],
+ "score": 1.0,
+ "content": "normalised text or phonemes in an end-to-end manner, resulting in models which",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 141,
+ 245,
+ 470,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 245,
+ 470,
+ 257
+ ],
+ "score": 1.0,
+ "content": "operate directly on character or phoneme input sequences and produce raw speech",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 255,
+ 470,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 255,
+ 470,
+ 268
+ ],
+ "score": 1.0,
+ "content": "audio outputs. Our proposed generator is feed-forward and thus efficient for both",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 266,
+ 470,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 266,
+ 470,
+ 279
+ ],
+ "score": 1.0,
+ "content": "training and inference, using a differentiable alignment scheme based on token",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 277,
+ 469,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 277,
+ 469,
+ 289
+ ],
+ "score": 1.0,
+ "content": "length prediction. It learns to produce high fidelity audio through a combination",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 288,
+ 470,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 288,
+ 470,
+ 301
+ ],
+ "score": 1.0,
+ "content": "of adversarial feedback and prediction losses constraining the generated audio to",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 299,
+ 470,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 299,
+ 470,
+ 312
+ ],
+ "score": 1.0,
+ "content": "roughly match the ground truth in terms of its total duration and mel-spectrogram.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 309,
+ 469,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 309,
+ 469,
+ 323
+ ],
+ "score": 1.0,
+ "content": "To allow the model to capture temporal variation in the generated audio, we employ",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 142,
+ 321,
+ 469,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 321,
+ 469,
+ 333
+ ],
+ "score": 1.0,
+ "content": "soft dynamic time warping in the spectrogram-based prediction loss. The resulting",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 332,
+ 470,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 332,
+ 470,
+ 345
+ ],
+ "score": 1.0,
+ "content": "model achieves a mean opinion score exceeding 4 on a 5 point scale, which",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 343,
+ 470,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 343,
+ 470,
+ 356
+ ],
+ "score": 1.0,
+ "content": "is comparable to the state-of-the-art models relying on multi-stage training and",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 142,
+ 354,
+ 239,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 354,
+ 239,
+ 366
+ ],
+ "score": 1.0,
+ "content": "additional supervision.1",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 388,
+ 206,
+ 401
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 208,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 208,
+ 404
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 413,
+ 505,
+ 512
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 505,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 505,
+ 426
+ ],
+ "score": 1.0,
+ "content": "A text-to-speech (TTS) system processes natural language text inputs to produce synthetic human-like",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 505,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 505,
+ 437
+ ],
+ "score": 1.0,
+ "content": "speech outputs. Typical TTS pipelines consist of a number of stages trained or designed independently",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 435,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 435,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "– e.g. text normalisation, aligned linguistic featurisation, mel-spectrogram synthesis, and raw audio",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 446,
+ 505,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 505,
+ 459
+ ],
+ "score": 1.0,
+ "content": "waveform synthesis (Taylor, 2009). Although these pipelines have proven capable of realistic and",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 471
+ ],
+ "score": 1.0,
+ "content": "high-fidelity speech synthesis and enjoy wide real-world use today, these modular approaches come",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 505,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 505,
+ 482
+ ],
+ "score": 1.0,
+ "content": "with a number of drawbacks. They often require supervision at each stage, in some cases necessitating",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 504,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 504,
+ 492
+ ],
+ "score": 1.0,
+ "content": "expensive “ground truth” annotations to guide the outputs of each stage, and sequential training of the",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 506,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 506,
+ 505
+ ],
+ "score": 1.0,
+ "content": "stages. Further, they are unable to reap the full potential rewards of data-driven “end-to-end\" learning",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 490,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 490,
+ 515
+ ],
+ "score": 1.0,
+ "content": "widely observed in a number of prediction and synthesis task domains across machine learning.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 515,
+ 505,
+ 593
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "In this work, we aim to simplify the TTS pipeline and take on the challenging task of synthesising",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 539
+ ],
+ "score": 1.0,
+ "content": "speech from text or phonemes in an end-to-end manner. We propose EATS – End-to-end Adversarial",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "Text-to-Speech – generative models for TTS trained adversarially (Goodfellow et al., 2014) that",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 549,
+ 505,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 549,
+ 505,
+ 561
+ ],
+ "score": 1.0,
+ "content": "operate on either pure text or raw (temporally unaligned) phoneme input sequences, and produce raw",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 559,
+ 506,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 559,
+ 506,
+ 573
+ ],
+ "score": 1.0,
+ "content": "speech waveforms as output. These models eliminate the typical intermediate bottlenecks present",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 571,
+ 505,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 505,
+ 583
+ ],
+ "score": 1.0,
+ "content": "in most state-of-the-art TTS engines by maintaining learnt intermediate feature representations",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 582,
+ 205,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 205,
+ 594
+ ],
+ "score": 1.0,
+ "content": "throughout the network.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 596,
+ 505,
+ 663
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 596,
+ 505,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 596,
+ 505,
+ 609
+ ],
+ "score": 1.0,
+ "content": "Our speech synthesis models are composed of two high-level submodules, detailed in Section 2. An",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 608,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 608,
+ 435,
+ 621
+ ],
+ "score": 1.0,
+ "content": "aligner processes the raw input sequence and produces relatively low-frequency",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 436,
+ 608,
+ 471,
+ 619
+ ],
+ "score": 0.73,
+ "content": "( 2 0 0 \\ : \\mathrm { H z } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 471,
+ 608,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "aligned",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 619,
+ 505,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 619,
+ 505,
+ 631
+ ],
+ "score": 1.0,
+ "content": "features in its own learnt, abstract feature space. The features output by the aligner may be thought",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 629,
+ 507,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 629,
+ 507,
+ 644
+ ],
+ "score": 1.0,
+ "content": "of as taking the place of the earlier stages of typical TTS pipelines – e.g., temporally aligned mel-",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 640,
+ 506,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 640,
+ 506,
+ 654
+ ],
+ "score": 1.0,
+ "content": "spectrograms or linguistic features. These features are then input to the decoder which upsamples the",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 651,
+ 437,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 651,
+ 332,
+ 664
+ ],
+ "score": 1.0,
+ "content": "features from the aligner by 1D convolutions to produce",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 332,
+ 652,
+ 363,
+ 662
+ ],
+ "score": 0.7,
+ "content": "2 4 \\mathrm { k H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 363,
+ 651,
+ 437,
+ 664
+ ],
+ "score": 1.0,
+ "content": "audio waveforms.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 39.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 666,
+ 504,
+ 689
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 679
+ ],
+ "score": 1.0,
+ "content": "By carefully designing the aligner and guiding training by a combination of adversarial feedback and",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 107,
+ 678,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 678,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "domain-specific loss functions, we demonstrate that a TTS system can be learnt nearly end-to-end,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43.5
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 699,
+ 329,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 699,
+ 329,
+ 712
+ ],
+ "score": 1.0,
+ "content": "∗Equal contribution. First author determined by coin toss.",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 128,
+ 710,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 710,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Listen to our model reading this abstract at: https://deepmind.com/research/",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 381,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 381,
+ 732
+ ],
+ "score": 1.0,
+ "content": "publications/End-to-End-Adversarial-Text-to-Speech",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 292,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 761
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 761
+ ],
+ "score": 1.0,
+ "content": "1",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 78,
+ 455,
+ 96
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 456,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 456,
+ 98
+ ],
+ "score": 1.0,
+ "content": "END-TO-END ADVERSARIAL TEXT-TO-SPEECH",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 113,
+ 112,
+ 479,
+ 146
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 112,
+ 480,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 112,
+ 480,
+ 126
+ ],
+ "score": 1.0,
+ "content": "Jeff Donahue∗, Sander Dieleman∗, Mikołaj Binkowski, Erich Elsen, Karen Simonyan ´ ∗",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 111,
+ 124,
+ 158,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 124,
+ 158,
+ 136
+ ],
+ "score": 1.0,
+ "content": "DeepMind",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 134,
+ 444,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 134,
+ 444,
+ 149
+ ],
+ "score": 1.0,
+ "content": "{jeffdonahue,sedielem,binek,eriche,simonyan}@google.com",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2,
+ "bbox_fs": [
+ 111,
+ 112,
+ 480,
+ 149
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 175,
+ 333,
+ 187
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 174,
+ 335,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 174,
+ 335,
+ 189
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 142,
+ 200,
+ 468,
+ 365
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 199,
+ 470,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 199,
+ 470,
+ 214
+ ],
+ "score": 1.0,
+ "content": "Modern text-to-speech synthesis pipelines typically involve multiple processing",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 212,
+ 470,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 212,
+ 470,
+ 223
+ ],
+ "score": 1.0,
+ "content": "stages, each of which is designed or learnt independently from the rest. In this",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 223,
+ 469,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 223,
+ 469,
+ 234
+ ],
+ "score": 1.0,
+ "content": "work, we take on the challenging task of learning to synthesise speech from",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 233,
+ 470,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 233,
+ 470,
+ 245
+ ],
+ "score": 1.0,
+ "content": "normalised text or phonemes in an end-to-end manner, resulting in models which",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 141,
+ 245,
+ 470,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 245,
+ 470,
+ 257
+ ],
+ "score": 1.0,
+ "content": "operate directly on character or phoneme input sequences and produce raw speech",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 255,
+ 470,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 255,
+ 470,
+ 268
+ ],
+ "score": 1.0,
+ "content": "audio outputs. Our proposed generator is feed-forward and thus efficient for both",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 266,
+ 470,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 266,
+ 470,
+ 279
+ ],
+ "score": 1.0,
+ "content": "training and inference, using a differentiable alignment scheme based on token",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 277,
+ 469,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 277,
+ 469,
+ 289
+ ],
+ "score": 1.0,
+ "content": "length prediction. It learns to produce high fidelity audio through a combination",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 288,
+ 470,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 288,
+ 470,
+ 301
+ ],
+ "score": 1.0,
+ "content": "of adversarial feedback and prediction losses constraining the generated audio to",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 299,
+ 470,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 299,
+ 470,
+ 312
+ ],
+ "score": 1.0,
+ "content": "roughly match the ground truth in terms of its total duration and mel-spectrogram.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 309,
+ 469,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 309,
+ 469,
+ 323
+ ],
+ "score": 1.0,
+ "content": "To allow the model to capture temporal variation in the generated audio, we employ",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 142,
+ 321,
+ 469,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 321,
+ 469,
+ 333
+ ],
+ "score": 1.0,
+ "content": "soft dynamic time warping in the spectrogram-based prediction loss. The resulting",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 332,
+ 470,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 332,
+ 470,
+ 345
+ ],
+ "score": 1.0,
+ "content": "model achieves a mean opinion score exceeding 4 on a 5 point scale, which",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 343,
+ 470,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 343,
+ 470,
+ 356
+ ],
+ "score": 1.0,
+ "content": "is comparable to the state-of-the-art models relying on multi-stage training and",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 142,
+ 354,
+ 239,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 354,
+ 239,
+ 366
+ ],
+ "score": 1.0,
+ "content": "additional supervision.1",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 141,
+ 199,
+ 470,
+ 366
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 388,
+ 206,
+ 401
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 208,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 208,
+ 404
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 413,
+ 505,
+ 512
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 505,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 505,
+ 426
+ ],
+ "score": 1.0,
+ "content": "A text-to-speech (TTS) system processes natural language text inputs to produce synthetic human-like",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 505,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 505,
+ 437
+ ],
+ "score": 1.0,
+ "content": "speech outputs. Typical TTS pipelines consist of a number of stages trained or designed independently",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 435,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 435,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "– e.g. text normalisation, aligned linguistic featurisation, mel-spectrogram synthesis, and raw audio",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 446,
+ 505,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 505,
+ 459
+ ],
+ "score": 1.0,
+ "content": "waveform synthesis (Taylor, 2009). Although these pipelines have proven capable of realistic and",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 506,
+ 471
+ ],
+ "score": 1.0,
+ "content": "high-fidelity speech synthesis and enjoy wide real-world use today, these modular approaches come",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 505,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 505,
+ 482
+ ],
+ "score": 1.0,
+ "content": "with a number of drawbacks. They often require supervision at each stage, in some cases necessitating",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 504,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 504,
+ 492
+ ],
+ "score": 1.0,
+ "content": "expensive “ground truth” annotations to guide the outputs of each stage, and sequential training of the",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 506,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 506,
+ 505
+ ],
+ "score": 1.0,
+ "content": "stages. Further, they are unable to reap the full potential rewards of data-driven “end-to-end\" learning",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 490,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 490,
+ 515
+ ],
+ "score": 1.0,
+ "content": "widely observed in a number of prediction and synthesis task domains across machine learning.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 104,
+ 413,
+ 506,
+ 515
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 515,
+ 505,
+ 593
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "In this work, we aim to simplify the TTS pipeline and take on the challenging task of synthesising",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 539
+ ],
+ "score": 1.0,
+ "content": "speech from text or phonemes in an end-to-end manner. We propose EATS – End-to-end Adversarial",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "Text-to-Speech – generative models for TTS trained adversarially (Goodfellow et al., 2014) that",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 549,
+ 505,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 549,
+ 505,
+ 561
+ ],
+ "score": 1.0,
+ "content": "operate on either pure text or raw (temporally unaligned) phoneme input sequences, and produce raw",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 559,
+ 506,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 559,
+ 506,
+ 573
+ ],
+ "score": 1.0,
+ "content": "speech waveforms as output. These models eliminate the typical intermediate bottlenecks present",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 571,
+ 505,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 505,
+ 583
+ ],
+ "score": 1.0,
+ "content": "in most state-of-the-art TTS engines by maintaining learnt intermediate feature representations",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 582,
+ 205,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 205,
+ 594
+ ],
+ "score": 1.0,
+ "content": "throughout the network.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 105,
+ 515,
+ 506,
+ 594
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 596,
+ 505,
+ 663
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 596,
+ 505,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 596,
+ 505,
+ 609
+ ],
+ "score": 1.0,
+ "content": "Our speech synthesis models are composed of two high-level submodules, detailed in Section 2. An",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 608,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 608,
+ 435,
+ 621
+ ],
+ "score": 1.0,
+ "content": "aligner processes the raw input sequence and produces relatively low-frequency",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 436,
+ 608,
+ 471,
+ 619
+ ],
+ "score": 0.73,
+ "content": "( 2 0 0 \\ : \\mathrm { H z } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 471,
+ 608,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "aligned",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 619,
+ 505,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 619,
+ 505,
+ 631
+ ],
+ "score": 1.0,
+ "content": "features in its own learnt, abstract feature space. The features output by the aligner may be thought",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 629,
+ 507,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 629,
+ 507,
+ 644
+ ],
+ "score": 1.0,
+ "content": "of as taking the place of the earlier stages of typical TTS pipelines – e.g., temporally aligned mel-",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 640,
+ 506,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 640,
+ 506,
+ 654
+ ],
+ "score": 1.0,
+ "content": "spectrograms or linguistic features. These features are then input to the decoder which upsamples the",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 651,
+ 437,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 651,
+ 332,
+ 664
+ ],
+ "score": 1.0,
+ "content": "features from the aligner by 1D convolutions to produce",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 332,
+ 652,
+ 363,
+ 662
+ ],
+ "score": 0.7,
+ "content": "2 4 \\mathrm { k H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 363,
+ 651,
+ 437,
+ 664
+ ],
+ "score": 1.0,
+ "content": "audio waveforms.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 39.5,
+ "bbox_fs": [
+ 105,
+ 596,
+ 507,
+ 664
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 666,
+ 504,
+ 689
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 679
+ ],
+ "score": 1.0,
+ "content": "By carefully designing the aligner and guiding training by a combination of adversarial feedback and",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 107,
+ 678,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 678,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "domain-specific loss functions, we demonstrate that a TTS system can be learnt nearly end-to-end,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "resulting in high-fidelity natural-sounding speech approaching the state-of-the-art TTS systems. Our",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 94,
+ 218,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 94,
+ 218,
+ 105
+ ],
+ "score": 1.0,
+ "content": "main contributions include:",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 43.5,
+ "bbox_fs": [
+ 106,
+ 666,
+ 505,
+ 689
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "resulting in high-fidelity natural-sounding speech approaching the state-of-the-art TTS systems. Our",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 94,
+ 218,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 94,
+ 218,
+ 105
+ ],
+ "score": 1.0,
+ "content": "main contributions include:",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 112,
+ 506,
+ 189
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 112,
+ 505,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 112,
+ 505,
+ 125
+ ],
+ "score": 1.0,
+ "content": "• A fully differentiable and efficient feed-forward aligner architecture that predicts the duration of",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 113,
+ 124,
+ 369,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 124,
+ 369,
+ 136
+ ],
+ "score": 1.0,
+ "content": "each input token and produces an audio-aligned representation.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 506,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 506,
+ 153
+ ],
+ "score": 1.0,
+ "content": "• The use of flexible dynamic time warping-based prediction losses to enforce alignment with input",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 114,
+ 151,
+ 482,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 151,
+ 482,
+ 163
+ ],
+ "score": 1.0,
+ "content": "conditioning while allowing the model to capture the variability of timing in human speech.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 506,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 506,
+ 180
+ ],
+ "score": 1.0,
+ "content": "• An overall system achieving a mean opinion score of 4.083, approaching the state of the art from",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 114,
+ 177,
+ 308,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 177,
+ 308,
+ 190
+ ],
+ "score": 1.0,
+ "content": "models trained using richer supervisory signals.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 203,
+ 173,
+ 216
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 174,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 174,
+ 219
+ ],
+ "score": 1.0,
+ "content": "2 METHOD",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 348
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 240
+ ],
+ "score": 1.0,
+ "content": "Our goal is to learn a neural network (the generator) which maps an input sequence of characters",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 237,
+ 506,
+ 251
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 226,
+ 251
+ ],
+ "score": 1.0,
+ "content": "or phonemes to raw audio at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 226,
+ 238,
+ 257,
+ 249
+ ],
+ "score": 0.67,
+ "content": "2 4 \\mathrm { k H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 237,
+ 506,
+ 251
+ ],
+ "score": 1.0,
+ "content": ". Beyond the vastly different lengths of the input and output",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 249,
+ 506,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 249,
+ 506,
+ 262
+ ],
+ "score": 1.0,
+ "content": "signals, this task is also challenging because the input and output are not aligned, i.e. it is not",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 260,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "known beforehand which output tokens each input token will correspond to. To address these",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 284
+ ],
+ "score": 1.0,
+ "content": "challenges, we divide the generator into two blocks: (i) the aligner, which maps the unaligned input",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 282,
+ 506,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 471,
+ 294
+ ],
+ "score": 1.0,
+ "content": "sequence to a representation which is aligned with the output, but has a lower sample rate of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 472,
+ 282,
+ 502,
+ 292
+ ],
+ "score": 0.69,
+ "content": "2 0 0 \\mathrm { H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 282,
+ 506,
+ 294
+ ],
+ "score": 1.0,
+ "content": ";",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 306
+ ],
+ "score": 1.0,
+ "content": "and (ii) the decoder, which upsamples the aligner’s output to the full audio frequency. The entire",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 506,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 506,
+ 316
+ ],
+ "score": 1.0,
+ "content": "generator architecture is differentiable, and is trained end to end. Importantly, it is also a feed-forward",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 315,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 315,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "convolutional network, which makes it well-suited for applications where fast batched inference",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 325,
+ 505,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 397,
+ 338
+ ],
+ "score": 1.0,
+ "content": "is important: our EATS implementation generates speech at a speed of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 326,
+ 421,
+ 336
+ ],
+ "score": 0.88,
+ "content": "2 0 0 \\times",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 325,
+ 505,
+ 338
+ ],
+ "score": 1.0,
+ "content": "realtime on a single",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 337,
+ 468,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 337,
+ 468,
+ 348
+ ],
+ "score": 1.0,
+ "content": "NVIDIA V100 GPU (see Appendix A and Table 3 for details). It is illustrated in Figure 1.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 451
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 506,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 506,
+ 364
+ ],
+ "score": 1.0,
+ "content": "The generator is inspired by GAN-TTS (Binkowski et al., 2020), a text-to-speech generative ad- ´",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 362,
+ 506,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 362,
+ 506,
+ 375
+ ],
+ "score": 1.0,
+ "content": "versarial network operating on aligned linguistic features. We employ the GAN-TTS generator",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 372,
+ 506,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 506,
+ 387
+ ],
+ "score": 1.0,
+ "content": "as the decoder in our model, but instead of upsampling pre-computed linguistic features, its input",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 398
+ ],
+ "score": 1.0,
+ "content": "comes from the aligner block. We make it speaker-conditional by feeding in a speaker embedding",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 395,
+ 222,
+ 408
+ ],
+ "score": 1.0,
+ "content": "s alongside the latent vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 398,
+ 228,
+ 406
+ ],
+ "score": 0.56,
+ "content": "\\mathbf { z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 229,
+ 395,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": ", to enable training on a larger dataset with recordings from multiple",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 419
+ ],
+ "score": 1.0,
+ "content": "speakers. We also adopt the multiple random window discriminators (RWDs) from GAN-TTS, which",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "have been proven effective for adversarial raw waveform modelling, and we preprocess real audio",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 428,
+ 506,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 428,
+ 215,
+ 441
+ ],
+ "score": 1.0,
+ "content": "input by applying a simple",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 216,
+ 429,
+ 223,
+ 440
+ ],
+ "score": 0.83,
+ "content": "\\mu",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 223,
+ 428,
+ 506,
+ 441
+ ],
+ "score": 1.0,
+ "content": "-law transform. Hence, the generator is trained to produce audio in the",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 107,
+ 438,
+ 446,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 441,
+ 114,
+ 451
+ ],
+ "score": 0.75,
+ "content": "\\mu",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 114,
+ 438,
+ 446,
+ 453
+ ],
+ "score": 1.0,
+ "content": "-law domain and we apply the inverse transformation to its outputs when sampling.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 454,
+ 348,
+ 466
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 348,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 348,
+ 467
+ ],
+ "score": 1.0,
+ "content": "The loss function we use to train the generator is as follows:",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 204,
+ 471,
+ 405,
+ 487
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 204,
+ 471,
+ 405,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 204,
+ 471,
+ 405,
+ 487
+ ],
+ "score": 0.89,
+ "content": "\\mathcal { L } _ { G } = \\mathcal { L } _ { G , \\mathrm { a d v } } + \\lambda _ { \\mathrm { p r e d } } \\cdot \\mathcal { L } _ { \\mathrm { p r e d } } ^ { \\prime \\prime } + \\lambda _ { \\mathrm { l e n g t h } } \\cdot \\mathcal { L } _ { \\mathrm { l e n g t h } } ,",
+ "type": "interline_equation",
+ "image_path": "27eac29c9a6d3953f508878899213ec3765f3d0e5217338928d3458581c3a01f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 30,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 204,
+ 471,
+ 405,
+ 487
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 491,
+ 505,
+ 579
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 506,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 134,
+ 505
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 492,
+ 164,
+ 503
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { L } _ { G , \\mathrm { a d v } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 164,
+ 490,
+ 506,
+ 505
+ ],
+ "score": 1.0,
+ "content": "is the adversarial loss, linear in the discriminators’ outputs, paired with the hinge",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 502,
+ 506,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 506,
+ 514
+ ],
+ "score": 1.0,
+ "content": "loss (Lim & Ye, 2017; Tran et al., 2017) used as the discriminators’ objective, as used in GAN-",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 511,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 511,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": "TTS (Binkowski et al., 2020). The use of an adversarial (Goodfellow et al., 2014) loss is an advantage ´",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 536
+ ],
+ "score": 1.0,
+ "content": "of our approach, as this setup allows for efficient feed-forward training and inference, and such",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 547
+ ],
+ "score": 1.0,
+ "content": "losses tend to be mode-seeking in practice, a useful behaviour in a strongly conditioned setting where",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 104,
+ 545,
+ 506,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 545,
+ 506,
+ 559
+ ],
+ "score": 1.0,
+ "content": "realism is an important design goal, as in the case of text-to-speech. In the remainder of this section,",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 101,
+ 552,
+ 509,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 101,
+ 552,
+ 353,
+ 580
+ ],
+ "score": 1.0,
+ "content": "we describe the aligner network and the auxiliary predictiondetail, and recap the components which were adopted from G",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 557,
+ 383,
+ 569
+ ],
+ "score": 0.88,
+ "content": "( \\mathcal { L } _ { \\mathrm { { p r e d } } } ^ { \\prime \\prime } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 552,
+ 431,
+ 580
+ ],
+ "score": 1.0,
+ "content": "and lengthS.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 431,
+ 558,
+ 466,
+ 569
+ ],
+ "score": 0.75,
+ "content": "( \\mathcal { L } _ { \\mathrm { l e n g t h } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 466,
+ 552,
+ 509,
+ 580
+ ],
+ "score": 1.0,
+ "content": "losses in",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 592,
+ 172,
+ 603
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 173,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 173,
+ 605
+ ],
+ "score": 1.0,
+ "content": "2.1 ALIGNER",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 506,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 208,
+ 624
+ ],
+ "score": 1.0,
+ "content": "Given a token sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 611,
+ 285,
+ 623
+ ],
+ "score": 0.91,
+ "content": "{ \\bf x } = ( x _ { 1 } , \\dots , x _ { N } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 285,
+ 610,
+ 327,
+ 624
+ ],
+ "score": 1.0,
+ "content": "of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 612,
+ 338,
+ 621
+ ],
+ "score": 0.8,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 610,
+ 506,
+ 624
+ ],
+ "score": 1.0,
+ "content": ", we first compute token representations",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 107,
+ 622,
+ 505,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 622,
+ 165,
+ 634
+ ],
+ "score": 0.92,
+ "content": "{ \\bf h } = f ( { \\bf x } , { \\bf z } , { \\bf s } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 166,
+ 622,
+ 197,
+ 634
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 197,
+ 622,
+ 204,
+ 633
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 622,
+ 505,
+ 634
+ ],
+ "score": 1.0,
+ "content": "is a stack of dilated convolutions (van den Oord et al., 2016) interspersed",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 447,
+ 645
+ ],
+ "score": 1.0,
+ "content": "with batch normalisation (Ioffe & Szegedy, 2015) and ReLU activations. The latents",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 447,
+ 635,
+ 454,
+ 643
+ ],
+ "score": 0.52,
+ "content": "\\mathbf { z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 455,
+ 632,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "and speaker",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": "embedding s modulate the scale and shift parameters of the batch normalisation layers (Dumoulin",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "et al., 2017; De Vries et al., 2017). We then predict the length for each input token individually:",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 107,
+ 666,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 666,
+ 172,
+ 678
+ ],
+ "score": 0.92,
+ "content": "l _ { n } = g ( h _ { n } , \\mathbf { z } , \\mathbf { s } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 666,
+ 204,
+ 678
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 667,
+ 211,
+ 677
+ ],
+ "score": 0.79,
+ "content": "g",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 212,
+ 666,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "is an MLP. We use a ReLU nonlinearity at the output to ensure that the",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 102,
+ 677,
+ 509,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 677,
+ 210,
+ 706
+ ],
+ "score": 1.0,
+ "content": "predicted lengths are non-sum of the token lengths:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 687,
+ 275,
+ 700
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { r } { e _ { n } = \\sum _ { m = 1 } ^ { n } l _ { m } } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 276,
+ 677,
+ 412,
+ 706
+ ],
+ "score": 1.0,
+ "content": "then find the predicted token end p, and the token centre positions as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 413,
+ 687,
+ 474,
+ 701
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r } { c _ { n } = e _ { n } - \\frac { 1 } { 2 } l _ { n } } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 474,
+ 677,
+ 509,
+ 706
+ ],
+ "score": 1.0,
+ "content": "mulative. Based",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "on these predicted positions, we can interpolate the token representations into an audio-aligned",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 104,
+ 709,
+ 507,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 709,
+ 173,
+ 723
+ ],
+ "score": 1.0,
+ "content": "representation at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 173,
+ 710,
+ 203,
+ 720
+ ],
+ "score": 0.52,
+ "content": "2 0 0 \\mathrm { H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 709,
+ 207,
+ 723
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 207,
+ 710,
+ 277,
+ 722
+ ],
+ "score": 0.86,
+ "content": "\\mathbf { a } = ( a _ { 1 } , \\ldots , a _ { S } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 709,
+ 307,
+ 723
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 307,
+ 709,
+ 349,
+ 722
+ ],
+ "score": 0.93,
+ "content": "S = \\lceil e _ { N } \\rceil",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 349,
+ 709,
+ 507,
+ 723
+ ],
+ "score": 1.0,
+ "content": "is the total number of output time steps.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 154,
+ 734
+ ],
+ "score": 1.0,
+ "content": "To compute",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 722,
+ 164,
+ 732
+ ],
+ "score": 0.84,
+ "content": "a _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 165,
+ 720,
+ 407,
+ 734
+ ],
+ "score": 1.0,
+ "content": ", we obtain interpolation weights for the token representations",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 721,
+ 420,
+ 732
+ ],
+ "score": 0.88,
+ "content": "h _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 420,
+ 720,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "using a softmax over",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 43.5
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 292,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "score": 1.0,
+ "content": "2",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 105
+ ],
+ "lines": [],
+ "index": 0.5,
+ "bbox_fs": [
+ 106,
+ 82,
+ 505,
+ 105
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 106,
+ 112,
+ 506,
+ 189
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 112,
+ 505,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 112,
+ 505,
+ 125
+ ],
+ "score": 1.0,
+ "content": "• A fully differentiable and efficient feed-forward aligner architecture that predicts the duration of",
+ "type": "text"
+ }
+ ],
+ "index": 2,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 113,
+ 124,
+ 369,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 124,
+ 369,
+ 136
+ ],
+ "score": 1.0,
+ "content": "each input token and produces an audio-aligned representation.",
+ "type": "text"
+ }
+ ],
+ "index": 3,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 506,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 506,
+ 153
+ ],
+ "score": 1.0,
+ "content": "• The use of flexible dynamic time warping-based prediction losses to enforce alignment with input",
+ "type": "text"
+ }
+ ],
+ "index": 4,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 151,
+ 482,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 151,
+ 482,
+ 163
+ ],
+ "score": 1.0,
+ "content": "conditioning while allowing the model to capture the variability of timing in human speech.",
+ "type": "text"
+ }
+ ],
+ "index": 5,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 506,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 506,
+ 180
+ ],
+ "score": 1.0,
+ "content": "• An overall system achieving a mean opinion score of 4.083, approaching the state of the art from",
+ "type": "text"
+ }
+ ],
+ "index": 6,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 114,
+ 177,
+ 308,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 177,
+ 308,
+ 190
+ ],
+ "score": 1.0,
+ "content": "models trained using richer supervisory signals.",
+ "type": "text"
+ }
+ ],
+ "index": 7,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 105,
+ 112,
+ 506,
+ 190
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 203,
+ 173,
+ 216
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 174,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 174,
+ 219
+ ],
+ "score": 1.0,
+ "content": "2 METHOD",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 348
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 240
+ ],
+ "score": 1.0,
+ "content": "Our goal is to learn a neural network (the generator) which maps an input sequence of characters",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 237,
+ 506,
+ 251
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 226,
+ 251
+ ],
+ "score": 1.0,
+ "content": "or phonemes to raw audio at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 226,
+ 238,
+ 257,
+ 249
+ ],
+ "score": 0.67,
+ "content": "2 4 \\mathrm { k H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 237,
+ 506,
+ 251
+ ],
+ "score": 1.0,
+ "content": ". Beyond the vastly different lengths of the input and output",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 249,
+ 506,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 249,
+ 506,
+ 262
+ ],
+ "score": 1.0,
+ "content": "signals, this task is also challenging because the input and output are not aligned, i.e. it is not",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 260,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "known beforehand which output tokens each input token will correspond to. To address these",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 284
+ ],
+ "score": 1.0,
+ "content": "challenges, we divide the generator into two blocks: (i) the aligner, which maps the unaligned input",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 282,
+ 506,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 471,
+ 294
+ ],
+ "score": 1.0,
+ "content": "sequence to a representation which is aligned with the output, but has a lower sample rate of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 472,
+ 282,
+ 502,
+ 292
+ ],
+ "score": 0.69,
+ "content": "2 0 0 \\mathrm { H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 282,
+ 506,
+ 294
+ ],
+ "score": 1.0,
+ "content": ";",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 306
+ ],
+ "score": 1.0,
+ "content": "and (ii) the decoder, which upsamples the aligner’s output to the full audio frequency. The entire",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 506,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 506,
+ 316
+ ],
+ "score": 1.0,
+ "content": "generator architecture is differentiable, and is trained end to end. Importantly, it is also a feed-forward",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 315,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 315,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "convolutional network, which makes it well-suited for applications where fast batched inference",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 325,
+ 505,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 397,
+ 338
+ ],
+ "score": 1.0,
+ "content": "is important: our EATS implementation generates speech at a speed of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 326,
+ 421,
+ 336
+ ],
+ "score": 0.88,
+ "content": "2 0 0 \\times",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 325,
+ 505,
+ 338
+ ],
+ "score": 1.0,
+ "content": "realtime on a single",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 337,
+ 468,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 337,
+ 468,
+ 348
+ ],
+ "score": 1.0,
+ "content": "NVIDIA V100 GPU (see Appendix A and Table 3 for details). It is illustrated in Figure 1.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 14,
+ "bbox_fs": [
+ 105,
+ 226,
+ 506,
+ 348
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 451
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 506,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 506,
+ 364
+ ],
+ "score": 1.0,
+ "content": "The generator is inspired by GAN-TTS (Binkowski et al., 2020), a text-to-speech generative ad- ´",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 362,
+ 506,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 362,
+ 506,
+ 375
+ ],
+ "score": 1.0,
+ "content": "versarial network operating on aligned linguistic features. We employ the GAN-TTS generator",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 372,
+ 506,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 506,
+ 387
+ ],
+ "score": 1.0,
+ "content": "as the decoder in our model, but instead of upsampling pre-computed linguistic features, its input",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 398
+ ],
+ "score": 1.0,
+ "content": "comes from the aligner block. We make it speaker-conditional by feeding in a speaker embedding",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 395,
+ 222,
+ 408
+ ],
+ "score": 1.0,
+ "content": "s alongside the latent vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 398,
+ 228,
+ 406
+ ],
+ "score": 0.56,
+ "content": "\\mathbf { z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 229,
+ 395,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": ", to enable training on a larger dataset with recordings from multiple",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 419
+ ],
+ "score": 1.0,
+ "content": "speakers. We also adopt the multiple random window discriminators (RWDs) from GAN-TTS, which",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "have been proven effective for adversarial raw waveform modelling, and we preprocess real audio",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 428,
+ 506,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 428,
+ 215,
+ 441
+ ],
+ "score": 1.0,
+ "content": "input by applying a simple",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 216,
+ 429,
+ 223,
+ 440
+ ],
+ "score": 0.83,
+ "content": "\\mu",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 223,
+ 428,
+ 506,
+ 441
+ ],
+ "score": 1.0,
+ "content": "-law transform. Hence, the generator is trained to produce audio in the",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 107,
+ 438,
+ 446,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 441,
+ 114,
+ 451
+ ],
+ "score": 0.75,
+ "content": "\\mu",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 114,
+ 438,
+ 446,
+ 453
+ ],
+ "score": 1.0,
+ "content": "-law domain and we apply the inverse transformation to its outputs when sampling.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 24,
+ "bbox_fs": [
+ 105,
+ 351,
+ 506,
+ 453
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 454,
+ 348,
+ 466
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 348,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 348,
+ 467
+ ],
+ "score": 1.0,
+ "content": "The loss function we use to train the generator is as follows:",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29,
+ "bbox_fs": [
+ 105,
+ 453,
+ 348,
+ 467
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 204,
+ 471,
+ 405,
+ 487
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 204,
+ 471,
+ 405,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 204,
+ 471,
+ 405,
+ 487
+ ],
+ "score": 0.89,
+ "content": "\\mathcal { L } _ { G } = \\mathcal { L } _ { G , \\mathrm { a d v } } + \\lambda _ { \\mathrm { p r e d } } \\cdot \\mathcal { L } _ { \\mathrm { p r e d } } ^ { \\prime \\prime } + \\lambda _ { \\mathrm { l e n g t h } } \\cdot \\mathcal { L } _ { \\mathrm { l e n g t h } } ,",
+ "type": "interline_equation",
+ "image_path": "27eac29c9a6d3953f508878899213ec3765f3d0e5217338928d3458581c3a01f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 30,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 204,
+ 471,
+ 405,
+ 487
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 491,
+ 505,
+ 579
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 506,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 134,
+ 505
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 492,
+ 164,
+ 503
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { L } _ { G , \\mathrm { a d v } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 164,
+ 490,
+ 506,
+ 505
+ ],
+ "score": 1.0,
+ "content": "is the adversarial loss, linear in the discriminators’ outputs, paired with the hinge",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 502,
+ 506,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 506,
+ 514
+ ],
+ "score": 1.0,
+ "content": "loss (Lim & Ye, 2017; Tran et al., 2017) used as the discriminators’ objective, as used in GAN-",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 511,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 511,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": "TTS (Binkowski et al., 2020). The use of an adversarial (Goodfellow et al., 2014) loss is an advantage ´",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 536
+ ],
+ "score": 1.0,
+ "content": "of our approach, as this setup allows for efficient feed-forward training and inference, and such",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 547
+ ],
+ "score": 1.0,
+ "content": "losses tend to be mode-seeking in practice, a useful behaviour in a strongly conditioned setting where",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 104,
+ 545,
+ 506,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 545,
+ 506,
+ 559
+ ],
+ "score": 1.0,
+ "content": "realism is an important design goal, as in the case of text-to-speech. In the remainder of this section,",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 101,
+ 552,
+ 509,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 101,
+ 552,
+ 353,
+ 580
+ ],
+ "score": 1.0,
+ "content": "we describe the aligner network and the auxiliary predictiondetail, and recap the components which were adopted from G",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 557,
+ 383,
+ 569
+ ],
+ "score": 0.88,
+ "content": "( \\mathcal { L } _ { \\mathrm { { p r e d } } } ^ { \\prime \\prime } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 552,
+ 431,
+ 580
+ ],
+ "score": 1.0,
+ "content": "and lengthS.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 431,
+ 558,
+ 466,
+ 569
+ ],
+ "score": 0.75,
+ "content": "( \\mathcal { L } _ { \\mathrm { l e n g t h } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 466,
+ 552,
+ 509,
+ 580
+ ],
+ "score": 1.0,
+ "content": "losses in",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 101,
+ 490,
+ 509,
+ 580
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 592,
+ 172,
+ 603
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 173,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 173,
+ 605
+ ],
+ "score": 1.0,
+ "content": "2.1 ALIGNER",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 506,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 208,
+ 624
+ ],
+ "score": 1.0,
+ "content": "Given a token sequence",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 611,
+ 285,
+ 623
+ ],
+ "score": 0.91,
+ "content": "{ \\bf x } = ( x _ { 1 } , \\dots , x _ { N } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 285,
+ 610,
+ 327,
+ 624
+ ],
+ "score": 1.0,
+ "content": "of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 612,
+ 338,
+ 621
+ ],
+ "score": 0.8,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 610,
+ 506,
+ 624
+ ],
+ "score": 1.0,
+ "content": ", we first compute token representations",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 107,
+ 622,
+ 505,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 622,
+ 165,
+ 634
+ ],
+ "score": 0.92,
+ "content": "{ \\bf h } = f ( { \\bf x } , { \\bf z } , { \\bf s } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 166,
+ 622,
+ 197,
+ 634
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 197,
+ 622,
+ 204,
+ 633
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 622,
+ 505,
+ 634
+ ],
+ "score": 1.0,
+ "content": "is a stack of dilated convolutions (van den Oord et al., 2016) interspersed",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 447,
+ 645
+ ],
+ "score": 1.0,
+ "content": "with batch normalisation (Ioffe & Szegedy, 2015) and ReLU activations. The latents",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 447,
+ 635,
+ 454,
+ 643
+ ],
+ "score": 0.52,
+ "content": "\\mathbf { z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 455,
+ 632,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "and speaker",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": "embedding s modulate the scale and shift parameters of the batch normalisation layers (Dumoulin",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "et al., 2017; De Vries et al., 2017). We then predict the length for each input token individually:",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 107,
+ 666,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 666,
+ 172,
+ 678
+ ],
+ "score": 0.92,
+ "content": "l _ { n } = g ( h _ { n } , \\mathbf { z } , \\mathbf { s } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 666,
+ 204,
+ 678
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 667,
+ 211,
+ 677
+ ],
+ "score": 0.79,
+ "content": "g",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 212,
+ 666,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "is an MLP. We use a ReLU nonlinearity at the output to ensure that the",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 102,
+ 677,
+ 509,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 677,
+ 210,
+ 706
+ ],
+ "score": 1.0,
+ "content": "predicted lengths are non-sum of the token lengths:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 687,
+ 275,
+ 700
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { r } { e _ { n } = \\sum _ { m = 1 } ^ { n } l _ { m } } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 276,
+ 677,
+ 412,
+ 706
+ ],
+ "score": 1.0,
+ "content": "then find the predicted token end p, and the token centre positions as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 413,
+ 687,
+ 474,
+ 701
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r } { c _ { n } = e _ { n } - \\frac { 1 } { 2 } l _ { n } } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 474,
+ 677,
+ 509,
+ 706
+ ],
+ "score": 1.0,
+ "content": "mulative. Based",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "on these predicted positions, we can interpolate the token representations into an audio-aligned",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 104,
+ 709,
+ 507,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 709,
+ 173,
+ 723
+ ],
+ "score": 1.0,
+ "content": "representation at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 173,
+ 710,
+ 203,
+ 720
+ ],
+ "score": 0.52,
+ "content": "2 0 0 \\mathrm { H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 709,
+ 207,
+ 723
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 207,
+ 710,
+ 277,
+ 722
+ ],
+ "score": 0.86,
+ "content": "\\mathbf { a } = ( a _ { 1 } , \\ldots , a _ { S } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 709,
+ 307,
+ 723
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 307,
+ 709,
+ 349,
+ 722
+ ],
+ "score": 0.93,
+ "content": "S = \\lceil e _ { N } \\rceil",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 349,
+ 709,
+ 507,
+ 723
+ ],
+ "score": 1.0,
+ "content": "is the total number of output time steps.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 154,
+ 734
+ ],
+ "score": 1.0,
+ "content": "To compute",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 722,
+ 164,
+ 732
+ ],
+ "score": 0.84,
+ "content": "a _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 165,
+ 720,
+ 407,
+ 734
+ ],
+ "score": 1.0,
+ "content": ", we obtain interpolation weights for the token representations",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 721,
+ 420,
+ 732
+ ],
+ "score": 0.88,
+ "content": "h _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 420,
+ 720,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "using a softmax over",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 226,
+ 95
+ ],
+ "score": 1.0,
+ "content": "the squared distance between",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 226,
+ 84,
+ 231,
+ 92
+ ],
+ "score": 0.76,
+ "content": "t",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 232,
+ 81,
+ 249,
+ 95
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 249,
+ 84,
+ 260,
+ 93
+ ],
+ "score": 0.86,
+ "content": "c _ { n }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 261,
+ 81,
+ 404,
+ 95
+ ],
+ "score": 1.0,
+ "content": ", scaled by a temperature parameter",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 405,
+ 82,
+ 416,
+ 92
+ ],
+ "score": 0.87,
+ "content": "\\sigma ^ { 2 }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 416,
+ 81,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": ", which we set to 10.0",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 204,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 204,
+ 106
+ ],
+ "score": 1.0,
+ "content": "(i.e. a Gaussian kernel):",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 43.5,
+ "bbox_fs": [
+ 102,
+ 610,
+ 509,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 226,
+ 95
+ ],
+ "score": 1.0,
+ "content": "the squared distance between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 226,
+ 84,
+ 231,
+ 92
+ ],
+ "score": 0.76,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 232,
+ 81,
+ 249,
+ 95
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 249,
+ 84,
+ 260,
+ 93
+ ],
+ "score": 0.86,
+ "content": "c _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 81,
+ 404,
+ 95
+ ],
+ "score": 1.0,
+ "content": ", scaled by a temperature parameter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 405,
+ 82,
+ 416,
+ 92
+ ],
+ "score": 0.87,
+ "content": "\\sigma ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 416,
+ 81,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": ", which we set to 10.0",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 204,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 204,
+ 106
+ ],
+ "score": 1.0,
+ "content": "(i.e. a Gaussian kernel):",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 228,
+ 106,
+ 383,
+ 138
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 228,
+ 106,
+ 383,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 228,
+ 106,
+ 383,
+ 138
+ ],
+ "score": 0.95,
+ "content": "w _ { t } ^ { n } = \\frac { \\exp { \\left( - \\sigma ^ { - 2 } ( t - c _ { n } ) ^ { 2 } \\right) } } { \\sum _ { m = 1 } ^ { N } \\exp { \\left( - \\sigma ^ { - 2 } ( t - c _ { m } ) ^ { 2 } \\right) } } .",
+ "type": "interline_equation",
+ "image_path": "45815cf2671e815e34f001fd64acc004229a13ba08900675bb14c40609d9b470.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 228,
+ 106,
+ 383,
+ 122.0
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 228,
+ 122.0,
+ 383,
+ 138.0
+ ],
+ "spans": [],
+ "index": 3
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 143,
+ 505,
+ 223
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 102,
+ 140,
+ 509,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 140,
+ 289,
+ 165
+ ],
+ "score": 1.0,
+ "content": "Using these weights, we can then compute",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 143,
+ 368,
+ 159
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r } { a _ { t } \\ = \\ \\sum _ { n = 1 } ^ { N } w _ { t } ^ { n } h _ { n } } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 368,
+ 140,
+ 509,
+ 165
+ ],
+ "score": 1.0,
+ "content": ", which amounts to non-uniform",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 156,
+ 506,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 506,
+ 171
+ ],
+ "score": 1.0,
+ "content": "interpolation. By predicting token lengths and obtaining positions using cumulative summation,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 168,
+ 505,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 168,
+ 505,
+ 180
+ ],
+ "score": 1.0,
+ "content": "instead of predicting positions directly, we implicitly enforce monotonicity of the alignment. Note",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 178,
+ 505,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 178,
+ 505,
+ 191
+ ],
+ "score": 1.0,
+ "content": "that tokens which have a non-monotonic effect on prosody, such as punctuation, can still affect the",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 190,
+ 505,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 190,
+ 336,
+ 202
+ ],
+ "score": 1.0,
+ "content": "entire utterance thanks to the stack of dilated convolutions",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 337,
+ 190,
+ 344,
+ 201
+ ],
+ "score": 0.84,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 344,
+ 190,
+ 505,
+ 202
+ ],
+ "score": 1.0,
+ "content": ", whose receptive field is large enough to",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 213
+ ],
+ "score": 1.0,
+ "content": "allow for propagation of information across the entire token sequence. The convolutions also ensure",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 211,
+ 499,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 499,
+ 225
+ ],
+ "score": 1.0,
+ "content": "generalisation across different sequence lengths. Appendix B includes pseudocode for the aligner.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 237,
+ 282,
+ 248
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 236,
+ 283,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 236,
+ 283,
+ 249
+ ],
+ "score": 1.0,
+ "content": "2.2 WINDOWED GENERATOR TRAINING",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 257,
+ 297,
+ 476
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 257,
+ 297,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 257,
+ 297,
+ 269
+ ],
+ "score": 1.0,
+ "content": "Training examples vary widely in length, from",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 267,
+ 298,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 267,
+ 298,
+ 280
+ ],
+ "score": 1.0,
+ "content": "about 1 to 20 seconds. We cannot pad all se-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 279,
+ 297,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 279,
+ 297,
+ 291
+ ],
+ "score": 1.0,
+ "content": "quences to a maximal length during training, as",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 289,
+ 298,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 298,
+ 302
+ ],
+ "score": 1.0,
+ "content": "this would be wasteful and prohibitively expen-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 300,
+ 297,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 218,
+ 313
+ ],
+ "score": 1.0,
+ "content": "sive: 20 seconds of audio at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 300,
+ 250,
+ 311
+ ],
+ "score": 0.67,
+ "content": "2 4 \\mathrm { k H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 300,
+ 297,
+ 313
+ ],
+ "score": 1.0,
+ "content": "correspond",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 311,
+ 298,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 298,
+ 323
+ ],
+ "score": 1.0,
+ "content": "to 480,000 timesteps, which results in high mem-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 323,
+ 297,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 323,
+ 297,
+ 334
+ ],
+ "score": 1.0,
+ "content": "ory requirements. Instead, we randomly extract",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 333,
+ 297,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 297,
+ 344
+ ],
+ "score": 1.0,
+ "content": "a 2 second window from each example, which",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 344,
+ 298,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 344,
+ 298,
+ 356
+ ],
+ "score": 1.0,
+ "content": "we will refer to as a training window, by uni-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 356,
+ 298,
+ 368
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 356,
+ 238,
+ 368
+ ],
+ "score": 1.0,
+ "content": "formly sampling a random offset",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 239,
+ 357,
+ 245,
+ 367
+ ],
+ "score": 0.72,
+ "content": "\\eta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 356,
+ 298,
+ 368
+ ],
+ "score": 1.0,
+ "content": ". The aligner",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 366,
+ 297,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 367,
+ 151,
+ 378
+ ],
+ "score": 1.0,
+ "content": "produces a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 151,
+ 366,
+ 182,
+ 377
+ ],
+ "score": 0.72,
+ "content": "2 0 0 \\mathrm { H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 367,
+ 297,
+ 378
+ ],
+ "score": 1.0,
+ "content": "audio-aligned representation",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 377,
+ 298,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 377,
+ 298,
+ 389
+ ],
+ "score": 1.0,
+ "content": "for this window, which is then fed to the de-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 389,
+ 297,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 297,
+ 400
+ ],
+ "score": 1.0,
+ "content": "coder (see Figure 1). Note that we only need",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 400,
+ 297,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 400,
+ 155,
+ 410
+ ],
+ "score": 1.0,
+ "content": "to compute",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 401,
+ 166,
+ 410
+ ],
+ "score": 0.85,
+ "content": "a _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 166,
+ 400,
+ 226,
+ 410
+ ],
+ "score": 1.0,
+ "content": "for time steps",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 227,
+ 400,
+ 232,
+ 409
+ ],
+ "score": 0.68,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 400,
+ 297,
+ 410
+ ],
+ "score": 1.0,
+ "content": "that fall within",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 410,
+ 298,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 298,
+ 422
+ ],
+ "score": 1.0,
+ "content": "the sampled window, but we do have to com-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 421,
+ 297,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 421,
+ 233,
+ 433
+ ],
+ "score": 1.0,
+ "content": "pute the predicted token lengths",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 233,
+ 421,
+ 243,
+ 432
+ ],
+ "score": 0.87,
+ "content": "l _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 243,
+ 421,
+ 297,
+ 433
+ ],
+ "score": 1.0,
+ "content": "for the entire",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 433,
+ 297,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 433,
+ 297,
+ 444
+ ],
+ "score": 1.0,
+ "content": "input sequence. During evaluation, we simply",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 443,
+ 297,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 443,
+ 297,
+ 455
+ ],
+ "score": 1.0,
+ "content": "produce the audio-aligned representation for the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 454,
+ 297,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 454,
+ 297,
+ 465
+ ],
+ "score": 1.0,
+ "content": "full utterance and run the decoder on it, which",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 282,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 282,
+ 477
+ ],
+ "score": 1.0,
+ "content": "is possible because it is fully convolutional.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 21.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 109,
+ 491,
+ 270,
+ 501
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 490,
+ 272,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 490,
+ 272,
+ 502
+ ],
+ "score": 1.0,
+ "content": "2.3 ADVERSARIAL DISCRIMINATORS",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "index": 53
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 306,
+ 232,
+ 504,
+ 525
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 306,
+ 232,
+ 504,
+ 525
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 232,
+ 504,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 232,
+ 504,
+ 525
+ ],
+ "score": 0.971,
+ "type": "image",
+ "image_path": "e8346596cd53edabed8d93e517d344c6bdce1b87d2b7209f1d540d66775d362c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 43.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 306,
+ 232,
+ 504,
+ 244.20833333333334
+ ],
+ "spans": [],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 306,
+ 244.20833333333334,
+ 504,
+ 256.4166666666667
+ ],
+ "spans": [],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 306,
+ 256.4166666666667,
+ 504,
+ 268.625
+ ],
+ "spans": [],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 306,
+ 268.625,
+ 504,
+ 280.8333333333333
+ ],
+ "spans": [],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 306,
+ 280.8333333333333,
+ 504,
+ 293.04166666666663
+ ],
+ "spans": [],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 306,
+ 293.04166666666663,
+ 504,
+ 305.24999999999994
+ ],
+ "spans": [],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 306,
+ 305.24999999999994,
+ 504,
+ 317.45833333333326
+ ],
+ "spans": [],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 306,
+ 317.45833333333326,
+ 504,
+ 329.6666666666666
+ ],
+ "spans": [],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 306,
+ 329.6666666666666,
+ 504,
+ 341.8749999999999
+ ],
+ "spans": [],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 306,
+ 341.8749999999999,
+ 504,
+ 354.0833333333332
+ ],
+ "spans": [],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 306,
+ 354.0833333333332,
+ 504,
+ 366.2916666666665
+ ],
+ "spans": [],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 306,
+ 366.2916666666665,
+ 504,
+ 378.49999999999983
+ ],
+ "spans": [],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 306,
+ 378.49999999999983,
+ 504,
+ 390.70833333333314
+ ],
+ "spans": [],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 306,
+ 390.70833333333314,
+ 504,
+ 402.91666666666646
+ ],
+ "spans": [],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 306,
+ 402.91666666666646,
+ 504,
+ 415.1249999999998
+ ],
+ "spans": [],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 306,
+ 415.1249999999998,
+ 504,
+ 427.3333333333331
+ ],
+ "spans": [],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 306,
+ 427.3333333333331,
+ 504,
+ 439.5416666666664
+ ],
+ "spans": [],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 306,
+ 439.5416666666664,
+ 504,
+ 451.7499999999997
+ ],
+ "spans": [],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 306,
+ 451.7499999999997,
+ 504,
+ 463.95833333333303
+ ],
+ "spans": [],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 306,
+ 463.95833333333303,
+ 504,
+ 476.16666666666634
+ ],
+ "spans": [],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 306,
+ 476.16666666666634,
+ 504,
+ 488.37499999999966
+ ],
+ "spans": [],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 306,
+ 488.37499999999966,
+ 504,
+ 500.583333333333
+ ],
+ "spans": [],
+ "index": 68
+ },
+ {
+ "bbox": [
+ 306,
+ 500.583333333333,
+ 504,
+ 512.7916666666663
+ ],
+ "spans": [],
+ "index": 69
+ },
+ {
+ "bbox": [
+ 306,
+ 512.7916666666663,
+ 504,
+ 524.9999999999997
+ ],
+ "spans": [],
+ "index": 70
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 305,
+ 531,
+ 505,
+ 663
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 531,
+ 505,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 531,
+ 505,
+ 543
+ ],
+ "score": 1.0,
+ "content": "Figure 1: A diagram of the generator, including the",
+ "type": "text"
+ }
+ ],
+ "index": 71
+ },
+ {
+ "bbox": [
+ 304,
+ 542,
+ 505,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 542,
+ 466,
+ 554
+ ],
+ "score": 1.0,
+ "content": "monotonic interpolation-based aligner.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 544,
+ 474,
+ 552
+ ],
+ "score": 0.74,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 474,
+ 542,
+ 505,
+ 554
+ ],
+ "score": 1.0,
+ "content": "and ch",
+ "type": "text"
+ }
+ ],
+ "index": 72
+ },
+ {
+ "bbox": [
+ 303,
+ 552,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 552,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "denote the latent Gaussian vector and the number",
+ "type": "text"
+ }
+ ],
+ "index": 73
+ },
+ {
+ "bbox": [
+ 303,
+ 564,
+ 506,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 564,
+ 506,
+ 577
+ ],
+ "score": 1.0,
+ "content": "of output channels, respectively. During training,",
+ "type": "text"
+ }
+ ],
+ "index": 74
+ },
+ {
+ "bbox": [
+ 304,
+ 575,
+ 505,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 575,
+ 505,
+ 586
+ ],
+ "score": 1.0,
+ "content": "audio windows have a fixed length of 2 seconds",
+ "type": "text"
+ }
+ ],
+ "index": 75
+ },
+ {
+ "bbox": [
+ 303,
+ 585,
+ 506,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 585,
+ 506,
+ 599
+ ],
+ "score": 1.0,
+ "content": "and are generated from the conditioning text using",
+ "type": "text"
+ }
+ ],
+ "index": 76
+ },
+ {
+ "bbox": [
+ 303,
+ 596,
+ 506,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 596,
+ 366,
+ 609
+ ],
+ "score": 1.0,
+ "content": "random offsets",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 367,
+ 599,
+ 374,
+ 609
+ ],
+ "score": 0.8,
+ "content": "\\eta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 374,
+ 596,
+ 506,
+ 609
+ ],
+ "score": 1.0,
+ "content": "and predicted phoneme lengths;",
+ "type": "text"
+ }
+ ],
+ "index": 77
+ },
+ {
+ "bbox": [
+ 303,
+ 608,
+ 505,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 608,
+ 505,
+ 619
+ ],
+ "score": 1.0,
+ "content": "the shaded areas in the logits grid and waveform",
+ "type": "text"
+ }
+ ],
+ "index": 78
+ },
+ {
+ "bbox": [
+ 303,
+ 619,
+ 506,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 619,
+ 506,
+ 631
+ ],
+ "score": 1.0,
+ "content": "are not synthesised. For inference (sampling), we",
+ "type": "text"
+ }
+ ],
+ "index": 79
+ },
+ {
+ "bbox": [
+ 303,
+ 630,
+ 506,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 630,
+ 317,
+ 642
+ ],
+ "score": 1.0,
+ "content": "set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 317,
+ 630,
+ 342,
+ 641
+ ],
+ "score": 0.91,
+ "content": "\\eta = 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 630,
+ 506,
+ 642
+ ],
+ "score": 1.0,
+ "content": ". In the No Phonemes ablation, the phone-",
+ "type": "text"
+ }
+ ],
+ "index": 80
+ },
+ {
+ "bbox": [
+ 303,
+ 641,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 641,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "mizer is skipped and the character sequence is fed",
+ "type": "text"
+ }
+ ],
+ "index": 81
+ },
+ {
+ "bbox": [
+ 304,
+ 651,
+ 402,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 651,
+ 402,
+ 664
+ ],
+ "score": 1.0,
+ "content": "directly into the aligner.",
+ "type": "text"
+ }
+ ],
+ "index": 82
+ }
+ ],
+ "index": 76.5
+ }
+ ],
+ "index": 60.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 510,
+ 297,
+ 674
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 510,
+ 297,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 510,
+ 297,
+ 522
+ ],
+ "score": 1.0,
+ "content": "Random window discriminators. We use an",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 106,
+ 521,
+ 297,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 297,
+ 532
+ ],
+ "score": 1.0,
+ "content": "ensemble of random window discriminators",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 106,
+ 532,
+ 297,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 532,
+ 297,
+ 543
+ ],
+ "score": 1.0,
+ "content": "(RWDs) adopted from GAN-TTS. Each RWD",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 105,
+ 542,
+ 298,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 298,
+ 556
+ ],
+ "score": 1.0,
+ "content": "operates on audio fragments of different lengths,",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 106,
+ 554,
+ 299,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 554,
+ 299,
+ 565
+ ],
+ "score": 1.0,
+ "content": "randomly sampled from the training window.",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 299,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 299,
+ 577
+ ],
+ "score": 1.0,
+ "content": "We use five RWDs with window sizes 240, 480,",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 106,
+ 576,
+ 297,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 297,
+ 586
+ ],
+ "score": 1.0,
+ "content": "960, 1920 and 3600. This enables each RWD to",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 298,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 298,
+ 598
+ ],
+ "score": 1.0,
+ "content": "operate at a different resolution. Note that 3600",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ },
+ {
+ "bbox": [
+ 105,
+ 598,
+ 298,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 153,
+ 609
+ ],
+ "score": 1.0,
+ "content": "samples at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 598,
+ 186,
+ 608
+ ],
+ "score": 0.74,
+ "content": "2 4 ~ \\mathrm { k H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 598,
+ 252,
+ 609
+ ],
+ "score": 1.0,
+ "content": "corresponds to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 598,
+ 284,
+ 609
+ ],
+ "score": 0.48,
+ "content": "1 5 0 ~ \\mathrm { m s }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 285,
+ 598,
+ 298,
+ 609
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 106,
+ 609,
+ 298,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 609,
+ 298,
+ 620
+ ],
+ "score": 1.0,
+ "content": "audio, so all RWDs operate on short timescales.",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ },
+ {
+ "bbox": [
+ 106,
+ 619,
+ 297,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 619,
+ 297,
+ 631
+ ],
+ "score": 1.0,
+ "content": "All RWDs in our model are unconditional with",
+ "type": "text"
+ }
+ ],
+ "index": 64
+ },
+ {
+ "bbox": [
+ 106,
+ 631,
+ 298,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 631,
+ 298,
+ 642
+ ],
+ "score": 1.0,
+ "content": "respect to text: they cannot access the text se-",
+ "type": "text"
+ }
+ ],
+ "index": 65
+ },
+ {
+ "bbox": [
+ 106,
+ 642,
+ 297,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 642,
+ 297,
+ 653
+ ],
+ "score": 1.0,
+ "content": "quence or the aligner output. (GAN-TTS uses",
+ "type": "text"
+ }
+ ],
+ "index": 66
+ },
+ {
+ "bbox": [
+ 106,
+ 652,
+ 297,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 652,
+ 297,
+ 664
+ ],
+ "score": 1.0,
+ "content": "10 RWDs, including 5 conditioned on linguistic",
+ "type": "text"
+ }
+ ],
+ "index": 67
+ },
+ {
+ "bbox": [
+ 105,
+ 662,
+ 298,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 662,
+ 298,
+ 677
+ ],
+ "score": 1.0,
+ "content": "features which we omit.) They are, however,",
+ "type": "text"
+ }
+ ],
+ "index": 83
+ }
+ ],
+ "index": 61
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 675,
+ 432,
+ 686
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 673,
+ 433,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 673,
+ 433,
+ 687
+ ],
+ "score": 1.0,
+ "content": "conditioned on the speaker, via projection embedding (Miyato & Koyama, 2018).",
+ "type": "text"
+ }
+ ],
+ "index": 84
+ }
+ ],
+ "index": 84
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "Spectrogram discriminator. We use an additional discriminator which operates on the full training",
+ "type": "text"
+ }
+ ],
+ "index": 85
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "window in the spectrogram domain. We extract log-scaled mel-spectrograms from the audio signals",
+ "type": "text"
+ }
+ ],
+ "index": 86
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": "and use the BigGAN-deep architecture (Brock et al., 2018), essentially treating the spectrograms as",
+ "type": "text"
+ }
+ ],
+ "index": 87
+ }
+ ],
+ "index": 86
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 26,
+ 292,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "3",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 105
+ ],
+ "lines": [],
+ "index": 0.5,
+ "bbox_fs": [
+ 105,
+ 81,
+ 506,
+ 106
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 228,
+ 106,
+ 383,
+ 138
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 228,
+ 106,
+ 383,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 228,
+ 106,
+ 383,
+ 138
+ ],
+ "score": 0.95,
+ "content": "w _ { t } ^ { n } = \\frac { \\exp { \\left( - \\sigma ^ { - 2 } ( t - c _ { n } ) ^ { 2 } \\right) } } { \\sum _ { m = 1 } ^ { N } \\exp { \\left( - \\sigma ^ { - 2 } ( t - c _ { m } ) ^ { 2 } \\right) } } .",
+ "type": "interline_equation",
+ "image_path": "45815cf2671e815e34f001fd64acc004229a13ba08900675bb14c40609d9b470.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 228,
+ 106,
+ 383,
+ 122.0
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 228,
+ 122.0,
+ 383,
+ 138.0
+ ],
+ "spans": [],
+ "index": 3
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 143,
+ 505,
+ 223
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 102,
+ 140,
+ 509,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 140,
+ 289,
+ 165
+ ],
+ "score": 1.0,
+ "content": "Using these weights, we can then compute",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 143,
+ 368,
+ 159
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r } { a _ { t } \\ = \\ \\sum _ { n = 1 } ^ { N } w _ { t } ^ { n } h _ { n } } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 368,
+ 140,
+ 509,
+ 165
+ ],
+ "score": 1.0,
+ "content": ", which amounts to non-uniform",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 156,
+ 506,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 506,
+ 171
+ ],
+ "score": 1.0,
+ "content": "interpolation. By predicting token lengths and obtaining positions using cumulative summation,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 168,
+ 505,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 168,
+ 505,
+ 180
+ ],
+ "score": 1.0,
+ "content": "instead of predicting positions directly, we implicitly enforce monotonicity of the alignment. Note",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 178,
+ 505,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 178,
+ 505,
+ 191
+ ],
+ "score": 1.0,
+ "content": "that tokens which have a non-monotonic effect on prosody, such as punctuation, can still affect the",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 190,
+ 505,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 190,
+ 336,
+ 202
+ ],
+ "score": 1.0,
+ "content": "entire utterance thanks to the stack of dilated convolutions",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 337,
+ 190,
+ 344,
+ 201
+ ],
+ "score": 0.84,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 344,
+ 190,
+ 505,
+ 202
+ ],
+ "score": 1.0,
+ "content": ", whose receptive field is large enough to",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 213
+ ],
+ "score": 1.0,
+ "content": "allow for propagation of information across the entire token sequence. The convolutions also ensure",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 211,
+ 499,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 499,
+ 225
+ ],
+ "score": 1.0,
+ "content": "generalisation across different sequence lengths. Appendix B includes pseudocode for the aligner.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 7,
+ "bbox_fs": [
+ 102,
+ 140,
+ 509,
+ 225
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 237,
+ 282,
+ 248
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 236,
+ 283,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 236,
+ 283,
+ 249
+ ],
+ "score": 1.0,
+ "content": "2.2 WINDOWED GENERATOR TRAINING",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 257,
+ 297,
+ 476
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 257,
+ 297,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 257,
+ 297,
+ 269
+ ],
+ "score": 1.0,
+ "content": "Training examples vary widely in length, from",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 267,
+ 298,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 267,
+ 298,
+ 280
+ ],
+ "score": 1.0,
+ "content": "about 1 to 20 seconds. We cannot pad all se-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 279,
+ 297,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 279,
+ 297,
+ 291
+ ],
+ "score": 1.0,
+ "content": "quences to a maximal length during training, as",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 289,
+ 298,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 298,
+ 302
+ ],
+ "score": 1.0,
+ "content": "this would be wasteful and prohibitively expen-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 300,
+ 297,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 218,
+ 313
+ ],
+ "score": 1.0,
+ "content": "sive: 20 seconds of audio at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 300,
+ 250,
+ 311
+ ],
+ "score": 0.67,
+ "content": "2 4 \\mathrm { k H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 300,
+ 297,
+ 313
+ ],
+ "score": 1.0,
+ "content": "correspond",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 311,
+ 298,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 298,
+ 323
+ ],
+ "score": 1.0,
+ "content": "to 480,000 timesteps, which results in high mem-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 323,
+ 297,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 323,
+ 297,
+ 334
+ ],
+ "score": 1.0,
+ "content": "ory requirements. Instead, we randomly extract",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 333,
+ 297,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 297,
+ 344
+ ],
+ "score": 1.0,
+ "content": "a 2 second window from each example, which",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 344,
+ 298,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 344,
+ 298,
+ 356
+ ],
+ "score": 1.0,
+ "content": "we will refer to as a training window, by uni-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 356,
+ 298,
+ 368
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 356,
+ 238,
+ 368
+ ],
+ "score": 1.0,
+ "content": "formly sampling a random offset",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 239,
+ 357,
+ 245,
+ 367
+ ],
+ "score": 0.72,
+ "content": "\\eta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 356,
+ 298,
+ 368
+ ],
+ "score": 1.0,
+ "content": ". The aligner",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 366,
+ 297,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 367,
+ 151,
+ 378
+ ],
+ "score": 1.0,
+ "content": "produces a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 151,
+ 366,
+ 182,
+ 377
+ ],
+ "score": 0.72,
+ "content": "2 0 0 \\mathrm { H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 367,
+ 297,
+ 378
+ ],
+ "score": 1.0,
+ "content": "audio-aligned representation",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 377,
+ 298,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 377,
+ 298,
+ 389
+ ],
+ "score": 1.0,
+ "content": "for this window, which is then fed to the de-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 389,
+ 297,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 297,
+ 400
+ ],
+ "score": 1.0,
+ "content": "coder (see Figure 1). Note that we only need",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 400,
+ 297,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 400,
+ 155,
+ 410
+ ],
+ "score": 1.0,
+ "content": "to compute",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 401,
+ 166,
+ 410
+ ],
+ "score": 0.85,
+ "content": "a _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 166,
+ 400,
+ 226,
+ 410
+ ],
+ "score": 1.0,
+ "content": "for time steps",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 227,
+ 400,
+ 232,
+ 409
+ ],
+ "score": 0.68,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 400,
+ 297,
+ 410
+ ],
+ "score": 1.0,
+ "content": "that fall within",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 410,
+ 298,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 298,
+ 422
+ ],
+ "score": 1.0,
+ "content": "the sampled window, but we do have to com-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 421,
+ 297,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 421,
+ 233,
+ 433
+ ],
+ "score": 1.0,
+ "content": "pute the predicted token lengths",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 233,
+ 421,
+ 243,
+ 432
+ ],
+ "score": 0.87,
+ "content": "l _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 243,
+ 421,
+ 297,
+ 433
+ ],
+ "score": 1.0,
+ "content": "for the entire",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 433,
+ 297,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 433,
+ 297,
+ 444
+ ],
+ "score": 1.0,
+ "content": "input sequence. During evaluation, we simply",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 443,
+ 297,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 443,
+ 297,
+ 455
+ ],
+ "score": 1.0,
+ "content": "produce the audio-aligned representation for the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 454,
+ 297,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 454,
+ 297,
+ 465
+ ],
+ "score": 1.0,
+ "content": "full utterance and run the decoder on it, which",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 282,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 282,
+ 477
+ ],
+ "score": 1.0,
+ "content": "is possible because it is fully convolutional.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 21.5,
+ "bbox_fs": [
+ 105,
+ 257,
+ 298,
+ 477
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 109,
+ 491,
+ 270,
+ 501
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 490,
+ 272,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 490,
+ 272,
+ 502
+ ],
+ "score": 1.0,
+ "content": "2.3 ADVERSARIAL DISCRIMINATORS",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "index": 53
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 306,
+ 232,
+ 504,
+ 525
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 306,
+ 232,
+ 504,
+ 525
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 232,
+ 504,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 232,
+ 504,
+ 525
+ ],
+ "score": 0.971,
+ "type": "image",
+ "image_path": "e8346596cd53edabed8d93e517d344c6bdce1b87d2b7209f1d540d66775d362c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 43.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 306,
+ 232,
+ 504,
+ 244.20833333333334
+ ],
+ "spans": [],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 306,
+ 244.20833333333334,
+ 504,
+ 256.4166666666667
+ ],
+ "spans": [],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 306,
+ 256.4166666666667,
+ 504,
+ 268.625
+ ],
+ "spans": [],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 306,
+ 268.625,
+ 504,
+ 280.8333333333333
+ ],
+ "spans": [],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 306,
+ 280.8333333333333,
+ 504,
+ 293.04166666666663
+ ],
+ "spans": [],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 306,
+ 293.04166666666663,
+ 504,
+ 305.24999999999994
+ ],
+ "spans": [],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 306,
+ 305.24999999999994,
+ 504,
+ 317.45833333333326
+ ],
+ "spans": [],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 306,
+ 317.45833333333326,
+ 504,
+ 329.6666666666666
+ ],
+ "spans": [],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 306,
+ 329.6666666666666,
+ 504,
+ 341.8749999999999
+ ],
+ "spans": [],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 306,
+ 341.8749999999999,
+ 504,
+ 354.0833333333332
+ ],
+ "spans": [],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 306,
+ 354.0833333333332,
+ 504,
+ 366.2916666666665
+ ],
+ "spans": [],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 306,
+ 366.2916666666665,
+ 504,
+ 378.49999999999983
+ ],
+ "spans": [],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 306,
+ 378.49999999999983,
+ 504,
+ 390.70833333333314
+ ],
+ "spans": [],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 306,
+ 390.70833333333314,
+ 504,
+ 402.91666666666646
+ ],
+ "spans": [],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 306,
+ 402.91666666666646,
+ 504,
+ 415.1249999999998
+ ],
+ "spans": [],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 306,
+ 415.1249999999998,
+ 504,
+ 427.3333333333331
+ ],
+ "spans": [],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 306,
+ 427.3333333333331,
+ 504,
+ 439.5416666666664
+ ],
+ "spans": [],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 306,
+ 439.5416666666664,
+ 504,
+ 451.7499999999997
+ ],
+ "spans": [],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 306,
+ 451.7499999999997,
+ 504,
+ 463.95833333333303
+ ],
+ "spans": [],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 306,
+ 463.95833333333303,
+ 504,
+ 476.16666666666634
+ ],
+ "spans": [],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 306,
+ 476.16666666666634,
+ 504,
+ 488.37499999999966
+ ],
+ "spans": [],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 306,
+ 488.37499999999966,
+ 504,
+ 500.583333333333
+ ],
+ "spans": [],
+ "index": 68
+ },
+ {
+ "bbox": [
+ 306,
+ 500.583333333333,
+ 504,
+ 512.7916666666663
+ ],
+ "spans": [],
+ "index": 69
+ },
+ {
+ "bbox": [
+ 306,
+ 512.7916666666663,
+ 504,
+ 524.9999999999997
+ ],
+ "spans": [],
+ "index": 70
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 305,
+ 531,
+ 505,
+ 663
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 531,
+ 505,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 531,
+ 505,
+ 543
+ ],
+ "score": 1.0,
+ "content": "Figure 1: A diagram of the generator, including the",
+ "type": "text"
+ }
+ ],
+ "index": 71
+ },
+ {
+ "bbox": [
+ 304,
+ 542,
+ 505,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 542,
+ 466,
+ 554
+ ],
+ "score": 1.0,
+ "content": "monotonic interpolation-based aligner.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 544,
+ 474,
+ 552
+ ],
+ "score": 0.74,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 474,
+ 542,
+ 505,
+ 554
+ ],
+ "score": 1.0,
+ "content": "and ch",
+ "type": "text"
+ }
+ ],
+ "index": 72
+ },
+ {
+ "bbox": [
+ 303,
+ 552,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 552,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "denote the latent Gaussian vector and the number",
+ "type": "text"
+ }
+ ],
+ "index": 73
+ },
+ {
+ "bbox": [
+ 303,
+ 564,
+ 506,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 564,
+ 506,
+ 577
+ ],
+ "score": 1.0,
+ "content": "of output channels, respectively. During training,",
+ "type": "text"
+ }
+ ],
+ "index": 74
+ },
+ {
+ "bbox": [
+ 304,
+ 575,
+ 505,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 575,
+ 505,
+ 586
+ ],
+ "score": 1.0,
+ "content": "audio windows have a fixed length of 2 seconds",
+ "type": "text"
+ }
+ ],
+ "index": 75
+ },
+ {
+ "bbox": [
+ 303,
+ 585,
+ 506,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 585,
+ 506,
+ 599
+ ],
+ "score": 1.0,
+ "content": "and are generated from the conditioning text using",
+ "type": "text"
+ }
+ ],
+ "index": 76
+ },
+ {
+ "bbox": [
+ 303,
+ 596,
+ 506,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 596,
+ 366,
+ 609
+ ],
+ "score": 1.0,
+ "content": "random offsets",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 367,
+ 599,
+ 374,
+ 609
+ ],
+ "score": 0.8,
+ "content": "\\eta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 374,
+ 596,
+ 506,
+ 609
+ ],
+ "score": 1.0,
+ "content": "and predicted phoneme lengths;",
+ "type": "text"
+ }
+ ],
+ "index": 77
+ },
+ {
+ "bbox": [
+ 303,
+ 608,
+ 505,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 608,
+ 505,
+ 619
+ ],
+ "score": 1.0,
+ "content": "the shaded areas in the logits grid and waveform",
+ "type": "text"
+ }
+ ],
+ "index": 78
+ },
+ {
+ "bbox": [
+ 303,
+ 619,
+ 506,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 619,
+ 506,
+ 631
+ ],
+ "score": 1.0,
+ "content": "are not synthesised. For inference (sampling), we",
+ "type": "text"
+ }
+ ],
+ "index": 79
+ },
+ {
+ "bbox": [
+ 303,
+ 630,
+ 506,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 630,
+ 317,
+ 642
+ ],
+ "score": 1.0,
+ "content": "set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 317,
+ 630,
+ 342,
+ 641
+ ],
+ "score": 0.91,
+ "content": "\\eta = 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 630,
+ 506,
+ 642
+ ],
+ "score": 1.0,
+ "content": ". In the No Phonemes ablation, the phone-",
+ "type": "text"
+ }
+ ],
+ "index": 80
+ },
+ {
+ "bbox": [
+ 303,
+ 641,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 641,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "mizer is skipped and the character sequence is fed",
+ "type": "text"
+ }
+ ],
+ "index": 81
+ },
+ {
+ "bbox": [
+ 304,
+ 651,
+ 402,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 651,
+ 402,
+ 664
+ ],
+ "score": 1.0,
+ "content": "directly into the aligner.",
+ "type": "text"
+ }
+ ],
+ "index": 82
+ }
+ ],
+ "index": 76.5
+ }
+ ],
+ "index": 60.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 510,
+ 297,
+ 674
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 510,
+ 297,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 510,
+ 297,
+ 522
+ ],
+ "score": 1.0,
+ "content": "Random window discriminators. We use an",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 106,
+ 521,
+ 297,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 297,
+ 532
+ ],
+ "score": 1.0,
+ "content": "ensemble of random window discriminators",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 106,
+ 532,
+ 297,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 532,
+ 297,
+ 543
+ ],
+ "score": 1.0,
+ "content": "(RWDs) adopted from GAN-TTS. Each RWD",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 105,
+ 542,
+ 298,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 298,
+ 556
+ ],
+ "score": 1.0,
+ "content": "operates on audio fragments of different lengths,",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 106,
+ 554,
+ 299,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 554,
+ 299,
+ 565
+ ],
+ "score": 1.0,
+ "content": "randomly sampled from the training window.",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 299,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 299,
+ 577
+ ],
+ "score": 1.0,
+ "content": "We use five RWDs with window sizes 240, 480,",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 106,
+ 576,
+ 297,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 297,
+ 586
+ ],
+ "score": 1.0,
+ "content": "960, 1920 and 3600. This enables each RWD to",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 298,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 298,
+ 598
+ ],
+ "score": 1.0,
+ "content": "operate at a different resolution. Note that 3600",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ },
+ {
+ "bbox": [
+ 105,
+ 598,
+ 298,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 153,
+ 609
+ ],
+ "score": 1.0,
+ "content": "samples at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 598,
+ 186,
+ 608
+ ],
+ "score": 0.74,
+ "content": "2 4 ~ \\mathrm { k H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 598,
+ 252,
+ 609
+ ],
+ "score": 1.0,
+ "content": "corresponds to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 598,
+ 284,
+ 609
+ ],
+ "score": 0.48,
+ "content": "1 5 0 ~ \\mathrm { m s }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 285,
+ 598,
+ 298,
+ 609
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 106,
+ 609,
+ 298,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 609,
+ 298,
+ 620
+ ],
+ "score": 1.0,
+ "content": "audio, so all RWDs operate on short timescales.",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ },
+ {
+ "bbox": [
+ 106,
+ 619,
+ 297,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 619,
+ 297,
+ 631
+ ],
+ "score": 1.0,
+ "content": "All RWDs in our model are unconditional with",
+ "type": "text"
+ }
+ ],
+ "index": 64
+ },
+ {
+ "bbox": [
+ 106,
+ 631,
+ 298,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 631,
+ 298,
+ 642
+ ],
+ "score": 1.0,
+ "content": "respect to text: they cannot access the text se-",
+ "type": "text"
+ }
+ ],
+ "index": 65
+ },
+ {
+ "bbox": [
+ 106,
+ 642,
+ 297,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 642,
+ 297,
+ 653
+ ],
+ "score": 1.0,
+ "content": "quence or the aligner output. (GAN-TTS uses",
+ "type": "text"
+ }
+ ],
+ "index": 66
+ },
+ {
+ "bbox": [
+ 106,
+ 652,
+ 297,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 652,
+ 297,
+ 664
+ ],
+ "score": 1.0,
+ "content": "10 RWDs, including 5 conditioned on linguistic",
+ "type": "text"
+ }
+ ],
+ "index": 67
+ },
+ {
+ "bbox": [
+ 105,
+ 662,
+ 298,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 662,
+ 298,
+ 677
+ ],
+ "score": 1.0,
+ "content": "features which we omit.) They are, however,",
+ "type": "text"
+ }
+ ],
+ "index": 83
+ },
+ {
+ "bbox": [
+ 106,
+ 673,
+ 433,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 673,
+ 433,
+ 687
+ ],
+ "score": 1.0,
+ "content": "conditioned on the speaker, via projection embedding (Miyato & Koyama, 2018).",
+ "type": "text"
+ }
+ ],
+ "index": 84
+ }
+ ],
+ "index": 61,
+ "bbox_fs": [
+ 105,
+ 510,
+ 299,
+ 677
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 675,
+ 432,
+ 686
+ ],
+ "lines": [],
+ "index": 84,
+ "bbox_fs": [
+ 106,
+ 673,
+ 433,
+ 687
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "Spectrogram discriminator. We use an additional discriminator which operates on the full training",
+ "type": "text"
+ }
+ ],
+ "index": 85
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "window in the spectrogram domain. We extract log-scaled mel-spectrograms from the audio signals",
+ "type": "text"
+ }
+ ],
+ "index": 86
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": "and use the BigGAN-deep architecture (Brock et al., 2018), essentially treating the spectrograms as",
+ "type": "text"
+ }
+ ],
+ "index": 87
+ }
+ ],
+ "index": 86,
+ "bbox_fs": [
+ 105,
+ 697,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 507,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 507,
+ 96
+ ],
+ "score": 1.0,
+ "content": "images. The spectrogram discriminator also uses speaker identity through projection embedding.",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 436,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 436,
+ 106
+ ],
+ "score": 1.0,
+ "content": "Details on the spectrogram discriminator architecture are included in Appendix C.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 117,
+ 275,
+ 128
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 116,
+ 277,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 116,
+ 277,
+ 129
+ ],
+ "score": 1.0,
+ "content": "2.4 SPECTROGRAM PREDICTION LOSS",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 135,
+ 505,
+ 234
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 136,
+ 506,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 136,
+ 506,
+ 147
+ ],
+ "score": 1.0,
+ "content": "In preliminary experiments, we discovered that adversarial feedback is insufficient to learn alignment.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 146,
+ 505,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 146,
+ 505,
+ 158
+ ],
+ "score": 1.0,
+ "content": "At the start of training, the aligner does not produce an accurate alignment, so the information in the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 158,
+ 506,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 158,
+ 506,
+ 169
+ ],
+ "score": 1.0,
+ "content": "input tokens is incorrectly temporally distributed. This encourages the decoder to ignore the aligner",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 169,
+ 505,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 169,
+ 505,
+ 180
+ ],
+ "score": 1.0,
+ "content": "output. The unconditional discriminators provide no useful learning signal to correct this. If we want",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 180,
+ 505,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 180,
+ 505,
+ 192
+ ],
+ "score": 1.0,
+ "content": "to use conditional discriminators instead, we face a different problem: we do not have aligned ground",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 190,
+ 505,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 505,
+ 203
+ ],
+ "score": 1.0,
+ "content": "truth. Conditional discriminators also need an aligner module, which cannot function correctly at the",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 213
+ ],
+ "score": 1.0,
+ "content": "start of training, effectively turning them into unconditional discriminators. Although it should be",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 211,
+ 506,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 506,
+ 225
+ ],
+ "score": 1.0,
+ "content": "possible in theory to train the discriminators’ aligner modules adversarially, we find that this does not",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 223,
+ 271,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 271,
+ 235
+ ],
+ "score": 1.0,
+ "content": "work in practice, and training gets stuck.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 237,
+ 505,
+ 304
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 238,
+ 506,
+ 251
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 238,
+ 506,
+ 251
+ ],
+ "score": 1.0,
+ "content": "Instead, we propose to guide learning by using an explicit prediction loss in the spectrogram domain:",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 250,
+ 505,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 175,
+ 261
+ ],
+ "score": 1.0,
+ "content": "we minimise the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 175,
+ 250,
+ 187,
+ 260
+ ],
+ "score": 0.89,
+ "content": "L _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 188,
+ 250,
+ 505,
+ 261
+ ],
+ "score": 1.0,
+ "content": "loss between the log-scaled mel-spectrograms of the generator output, and the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 260,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "corresponding ground truth training window. This helps training to take off, and renders conditional",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 506,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 335,
+ 285
+ ],
+ "score": 1.0,
+ "content": "discriminators unnecessary, simplifying the model. Let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 335,
+ 271,
+ 354,
+ 283
+ ],
+ "score": 0.91,
+ "content": "S _ { \\mathrm { g e n } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 355,
+ 270,
+ 506,
+ 285
+ ],
+ "score": 1.0,
+ "content": "be the spectrogram of the generated",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 282,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 133,
+ 295
+ ],
+ "score": 1.0,
+ "content": "audio,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 282,
+ 149,
+ 294
+ ],
+ "score": 0.9,
+ "content": "S _ { \\mathrm { g t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 149,
+ 282,
+ 373,
+ 295
+ ],
+ "score": 1.0,
+ "content": "the spectrogram of the corresponding ground truth, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 374,
+ 282,
+ 401,
+ 294
+ ],
+ "score": 0.93,
+ "content": "S [ t , f ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 401,
+ 282,
+ 505,
+ 295
+ ],
+ "score": 1.0,
+ "content": "the log-scaled magnitude",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 293,
+ 376,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 293,
+ 155,
+ 306
+ ],
+ "score": 1.0,
+ "content": "at time step",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 294,
+ 160,
+ 303
+ ],
+ "score": 0.74,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 160,
+ 293,
+ 253,
+ 306
+ ],
+ "score": 1.0,
+ "content": "and mel-frequency bin",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 254,
+ 293,
+ 261,
+ 304
+ ],
+ "score": 0.83,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 293,
+ 376,
+ 306
+ ],
+ "score": 1.0,
+ "content": ". Then the prediction loss is:",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 209,
+ 315,
+ 403,
+ 333
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 209,
+ 315,
+ 403,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 209,
+ 315,
+ 403,
+ 333
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { p r e d } } = \\frac { 1 } { F } \\sum _ { t = 1 } ^ { T } \\sum _ { f = 1 } ^ { F } | S _ { \\mathrm { g e n } } [ t , f ] - S _ { \\mathrm { g t } } [ t , f ] | . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "dfef56075b01fa14853b720235db6f8454bc544cb2c4c329b0b4af381fd1c3ba.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 18,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 209,
+ 315,
+ 403,
+ 333
+ ],
+ "spans": [],
+ "index": 18
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 338,
+ 505,
+ 417
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 339,
+ 505,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 340,
+ 115,
+ 349
+ ],
+ "score": 0.8,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 115,
+ 339,
+ 134,
+ 352
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 340,
+ 144,
+ 349
+ ],
+ "score": 0.83,
+ "content": "F",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 144,
+ 339,
+ 505,
+ 352
+ ],
+ "score": 1.0,
+ "content": "are the total number of time steps and mel-frequency bins respectively. Computing the",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 363
+ ],
+ "score": 1.0,
+ "content": "prediction loss in the spectrogram domain, rather than the time domain, has the advantage of",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "increased invariance to phase differences between the generated and ground truth signals, which are",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 372,
+ 506,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 372,
+ 506,
+ 385
+ ],
+ "score": 1.0,
+ "content": "not perceptually salient. Seeing as the spectrogram extraction operation has several hyperparameters",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 384,
+ 504,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 384,
+ 504,
+ 395
+ ],
+ "score": 1.0,
+ "content": "and its implementation is not standardised, we provide the code we used for this in Appendix D. We",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 394,
+ 504,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 279,
+ 407
+ ],
+ "score": 1.0,
+ "content": "applied a small amount of jitter (by up to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 279,
+ 394,
+ 298,
+ 405
+ ],
+ "score": 0.86,
+ "content": "\\pm 6 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 299,
+ 394,
+ 345,
+ 407
+ ],
+ "score": 1.0,
+ "content": "samples at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 394,
+ 378,
+ 405
+ ],
+ "score": 0.72,
+ "content": "2 4 \\mathrm { k H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 378,
+ 394,
+ 504,
+ 407
+ ],
+ "score": 1.0,
+ "content": ") to the ground truth waveform",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 421,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 179,
+ 419
+ ],
+ "score": 1.0,
+ "content": "before computing",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 180,
+ 405,
+ 194,
+ 417
+ ],
+ "score": 0.9,
+ "content": "S _ { \\mathrm { g t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 195,
+ 405,
+ 421,
+ 419
+ ],
+ "score": 1.0,
+ "content": ", which helped to reduce artifacts in the generated audio.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 420,
+ 505,
+ 487
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 506,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 506,
+ 433
+ ],
+ "score": 1.0,
+ "content": "The inability to learn alignment from adversarial feedback alone is worth expanding on: likelihood-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 443
+ ],
+ "score": 1.0,
+ "content": "based autoregressive models have no issues learning alignment, because they are able to benefit",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 443,
+ 504,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 443,
+ 504,
+ 454
+ ],
+ "score": 1.0,
+ "content": "from teacher forcing (Williams & Zipser, 1989) during training: the model is trained to perform",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 466
+ ],
+ "score": 1.0,
+ "content": "next step prediction on each sequence step, given the preceding ground truth, and it is expected to",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 478
+ ],
+ "score": 1.0,
+ "content": "infer alignment only one step at a time. This is not compatible with feed-forward adversarial models",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 475,
+ 469,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 469,
+ 487
+ ],
+ "score": 1.0,
+ "content": "however, so the prediction loss is necessary to bootstrap alignment learning for our model.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 490,
+ 505,
+ 545
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 373,
+ 503
+ ],
+ "score": 1.0,
+ "content": "Note that although we make use of mel-spectrograms for training in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 374,
+ 490,
+ 398,
+ 502
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { L } _ { \\mathrm { p r e d } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 489,
+ 505,
+ 503
+ ],
+ "score": 1.0,
+ "content": "(and to compute the inputs",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 514
+ ],
+ "score": 1.0,
+ "content": "for the spectrogram discriminator, Section 2.3), the generator itself does not produce spectrograms",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 512,
+ 505,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 512,
+ 505,
+ 525
+ ],
+ "score": 1.0,
+ "content": "as part of the generation process. Rather, its outputs are raw waveforms, and we convert these",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 536
+ ],
+ "score": 1.0,
+ "content": "waveforms to spectrograms only for training (backpropagating gradients through the waveform to",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 535,
+ 267,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 535,
+ 267,
+ 547
+ ],
+ "score": 1.0,
+ "content": "mel-spectrogram conversion operation).",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 557,
+ 240,
+ 568
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 556,
+ 241,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 556,
+ 241,
+ 569
+ ],
+ "score": 1.0,
+ "content": "2.5 DYNAMIC TIME WARPING",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 575,
+ 505,
+ 642
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "The spectrogram prediction loss incorrectly assumes that token lengths are deterministic. We can",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "relax the requirement that the generated and ground truth spectrograms are exactly aligned, by",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "score": 1.0,
+ "content": "incorporating dynamic time warping (DTW) (Sakoe, 1971; Sakoe & Chiba, 1978). We calculate the",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 609,
+ 506,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 609,
+ 368,
+ 622
+ ],
+ "score": 1.0,
+ "content": "prediction loss by iteratively finding a minimal-cost alignment path",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 369,
+ 610,
+ 375,
+ 620
+ ],
+ "score": 0.82,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 609,
+ 506,
+ 622
+ ],
+ "score": 1.0,
+ "content": "between the generated and target",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 619,
+ 503,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 619,
+ 164,
+ 634
+ ],
+ "score": 1.0,
+ "content": "spectrograms,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 620,
+ 184,
+ 631
+ ],
+ "score": 0.9,
+ "content": "S _ { \\mathrm { g e n } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 619,
+ 202,
+ 634
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 203,
+ 620,
+ 217,
+ 632
+ ],
+ "score": 0.9,
+ "content": "S _ { \\mathrm { g t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 218,
+ 619,
+ 427,
+ 634
+ ],
+ "score": 1.0,
+ "content": ". We start at the first time step in both spectrograms:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 428,
+ 621,
+ 503,
+ 632
+ ],
+ "score": 0.92,
+ "content": "p _ { \\mathrm { g e n , 1 } } = p _ { \\mathrm { g t , 1 } } = 1",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 630,
+ 338,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 630,
+ 175,
+ 643
+ ],
+ "score": 1.0,
+ "content": "At each iteration",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 175,
+ 632,
+ 182,
+ 640
+ ],
+ "score": 0.81,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 630,
+ 338,
+ 643
+ ],
+ "score": 1.0,
+ "content": ", we take one of three possible actions:",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 130,
+ 649,
+ 493,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 647,
+ 491,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 647,
+ 266,
+ 664
+ ],
+ "score": 1.0,
+ "content": "1. go to the next time step in both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 266,
+ 649,
+ 491,
+ 662
+ ],
+ "score": 0.41,
+ "content": "S _ { \\mathrm { g e n } } , S _ { \\mathrm { g t } } \\colon p _ { \\mathrm { g e n } , k + 1 } = p _ { \\mathrm { g e n } , k } + 1 , p _ { \\mathrm { g t } , k + 1 } = p _ { \\mathrm { g t } , k } + 1 ;",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 128,
+ 663,
+ 452,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 663,
+ 248,
+ 678
+ ],
+ "score": 1.0,
+ "content": "2. go to the next time step in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 665,
+ 263,
+ 677
+ ],
+ "score": 0.89,
+ "content": "S _ { \\mathrm { g t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 663,
+ 286,
+ 678
+ ],
+ "score": 1.0,
+ "content": "only:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 287,
+ 665,
+ 448,
+ 677
+ ],
+ "score": 0.52,
+ "content": "p _ { \\mathrm { g e n } , k + 1 } = p _ { \\mathrm { g e n } , k } , p _ { \\mathrm { g t } , k + 1 } = p _ { \\mathrm { g t } , k } + 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 448,
+ 663,
+ 452,
+ 678
+ ],
+ "score": 1.0,
+ "content": ";",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 128,
+ 678,
+ 456,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 678,
+ 248,
+ 694
+ ],
+ "score": 1.0,
+ "content": "3. go to the next time step in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 249,
+ 680,
+ 268,
+ 692
+ ],
+ "score": 0.91,
+ "content": "S _ { \\mathrm { g e n } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 268,
+ 678,
+ 293,
+ 694
+ ],
+ "score": 1.0,
+ "content": "only:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 293,
+ 681,
+ 383,
+ 692
+ ],
+ "score": 0.77,
+ "content": "p _ { \\mathrm { g e n } , k + 1 } = p _ { \\mathrm { g e n } , k } + 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 678,
+ 456,
+ 694
+ ],
+ "score": 1.0,
+ "content": ", pgt,k+1 = pgt,k.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 45
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 698,
+ 503,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 196,
+ 713
+ ],
+ "score": 1.0,
+ "content": "The resulting path is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 699,
+ 372,
+ 712
+ ],
+ "score": 0.89,
+ "content": "p = \\langle ( p _ { \\mathrm { g e n } , 1 } , p _ { \\mathrm { g t } , 1 } ) , \\dots , ( p _ { \\mathrm { g e n } , K _ { p } } , p _ { \\mathrm { g t } , K _ { p } } ) \\rangle",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 372,
+ 697,
+ 405,
+ 713
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 699,
+ 420,
+ 711
+ ],
+ "score": 0.89,
+ "content": "K _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 420,
+ 697,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "is the length. Each",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 708,
+ 506,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 260,
+ 724
+ ],
+ "score": 1.0,
+ "content": "action is assigned a cost based on the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 260,
+ 711,
+ 272,
+ 721
+ ],
+ "score": 0.86,
+ "content": "L _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 708,
+ 346,
+ 724
+ ],
+ "score": 1.0,
+ "content": "distance between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 347,
+ 710,
+ 396,
+ 722
+ ],
+ "score": 0.91,
+ "content": "S _ { \\mathrm { g e n } } [ p _ { \\mathrm { g e n } , k } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 396,
+ 708,
+ 414,
+ 724
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 415,
+ 710,
+ 454,
+ 722
+ ],
+ "score": 0.92,
+ "content": "\\mathrm { \\dot { \\cal S } } _ { \\mathrm { g t } } [ p _ { \\mathrm { g t } , k } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 454,
+ 708,
+ 506,
+ 724
+ ],
+ "score": 1.0,
+ "content": ", and a warp",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 138,
+ 734
+ ],
+ "score": 1.0,
+ "content": "penalty",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 139,
+ 722,
+ 147,
+ 730
+ ],
+ "score": 0.71,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 147,
+ 720,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": "which is incurred if we choose not to advance both spectrograms in lockstep (i.e. we are",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 292,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 12,
+ "width": 8
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "list",
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 507,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 507,
+ 96
+ ],
+ "score": 1.0,
+ "content": "images. The spectrogram discriminator also uses speaker identity through projection embedding.",
+ "type": "text"
+ }
+ ],
+ "index": 0,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 436,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 436,
+ 106
+ ],
+ "score": 1.0,
+ "content": "Details on the spectrogram discriminator architecture are included in Appendix C.",
+ "type": "text"
+ }
+ ],
+ "index": 1,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 0.5,
+ "bbox_fs": [
+ 105,
+ 81,
+ 507,
+ 106
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 117,
+ 275,
+ 128
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 116,
+ 277,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 116,
+ 277,
+ 129
+ ],
+ "score": 1.0,
+ "content": "2.4 SPECTROGRAM PREDICTION LOSS",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 135,
+ 505,
+ 234
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 136,
+ 506,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 136,
+ 506,
+ 147
+ ],
+ "score": 1.0,
+ "content": "In preliminary experiments, we discovered that adversarial feedback is insufficient to learn alignment.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 146,
+ 505,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 146,
+ 505,
+ 158
+ ],
+ "score": 1.0,
+ "content": "At the start of training, the aligner does not produce an accurate alignment, so the information in the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 158,
+ 506,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 158,
+ 506,
+ 169
+ ],
+ "score": 1.0,
+ "content": "input tokens is incorrectly temporally distributed. This encourages the decoder to ignore the aligner",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 169,
+ 505,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 169,
+ 505,
+ 180
+ ],
+ "score": 1.0,
+ "content": "output. The unconditional discriminators provide no useful learning signal to correct this. If we want",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 180,
+ 505,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 180,
+ 505,
+ 192
+ ],
+ "score": 1.0,
+ "content": "to use conditional discriminators instead, we face a different problem: we do not have aligned ground",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 190,
+ 505,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 505,
+ 203
+ ],
+ "score": 1.0,
+ "content": "truth. Conditional discriminators also need an aligner module, which cannot function correctly at the",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 213
+ ],
+ "score": 1.0,
+ "content": "start of training, effectively turning them into unconditional discriminators. Although it should be",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 211,
+ 506,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 506,
+ 225
+ ],
+ "score": 1.0,
+ "content": "possible in theory to train the discriminators’ aligner modules adversarially, we find that this does not",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 223,
+ 271,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 271,
+ 235
+ ],
+ "score": 1.0,
+ "content": "work in practice, and training gets stuck.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 7,
+ "bbox_fs": [
+ 105,
+ 136,
+ 506,
+ 235
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 237,
+ 505,
+ 304
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 238,
+ 506,
+ 251
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 238,
+ 506,
+ 251
+ ],
+ "score": 1.0,
+ "content": "Instead, we propose to guide learning by using an explicit prediction loss in the spectrogram domain:",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 250,
+ 505,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 175,
+ 261
+ ],
+ "score": 1.0,
+ "content": "we minimise the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 175,
+ 250,
+ 187,
+ 260
+ ],
+ "score": 0.89,
+ "content": "L _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 188,
+ 250,
+ 505,
+ 261
+ ],
+ "score": 1.0,
+ "content": "loss between the log-scaled mel-spectrograms of the generator output, and the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 260,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "corresponding ground truth training window. This helps training to take off, and renders conditional",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 506,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 335,
+ 285
+ ],
+ "score": 1.0,
+ "content": "discriminators unnecessary, simplifying the model. Let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 335,
+ 271,
+ 354,
+ 283
+ ],
+ "score": 0.91,
+ "content": "S _ { \\mathrm { g e n } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 355,
+ 270,
+ 506,
+ 285
+ ],
+ "score": 1.0,
+ "content": "be the spectrogram of the generated",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 282,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 133,
+ 295
+ ],
+ "score": 1.0,
+ "content": "audio,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 282,
+ 149,
+ 294
+ ],
+ "score": 0.9,
+ "content": "S _ { \\mathrm { g t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 149,
+ 282,
+ 373,
+ 295
+ ],
+ "score": 1.0,
+ "content": "the spectrogram of the corresponding ground truth, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 374,
+ 282,
+ 401,
+ 294
+ ],
+ "score": 0.93,
+ "content": "S [ t , f ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 401,
+ 282,
+ 505,
+ 295
+ ],
+ "score": 1.0,
+ "content": "the log-scaled magnitude",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 293,
+ 376,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 293,
+ 155,
+ 306
+ ],
+ "score": 1.0,
+ "content": "at time step",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 294,
+ 160,
+ 303
+ ],
+ "score": 0.74,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 160,
+ 293,
+ 253,
+ 306
+ ],
+ "score": 1.0,
+ "content": "and mel-frequency bin",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 254,
+ 293,
+ 261,
+ 304
+ ],
+ "score": 0.83,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 293,
+ 376,
+ 306
+ ],
+ "score": 1.0,
+ "content": ". Then the prediction loss is:",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 105,
+ 238,
+ 506,
+ 306
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 209,
+ 315,
+ 403,
+ 333
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 209,
+ 315,
+ 403,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 209,
+ 315,
+ 403,
+ 333
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { p r e d } } = \\frac { 1 } { F } \\sum _ { t = 1 } ^ { T } \\sum _ { f = 1 } ^ { F } | S _ { \\mathrm { g e n } } [ t , f ] - S _ { \\mathrm { g t } } [ t , f ] | . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "dfef56075b01fa14853b720235db6f8454bc544cb2c4c329b0b4af381fd1c3ba.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 18,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 209,
+ 315,
+ 403,
+ 333
+ ],
+ "spans": [],
+ "index": 18
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 338,
+ 505,
+ 417
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 339,
+ 505,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 340,
+ 115,
+ 349
+ ],
+ "score": 0.8,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 115,
+ 339,
+ 134,
+ 352
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 340,
+ 144,
+ 349
+ ],
+ "score": 0.83,
+ "content": "F",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 144,
+ 339,
+ 505,
+ 352
+ ],
+ "score": 1.0,
+ "content": "are the total number of time steps and mel-frequency bins respectively. Computing the",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 506,
+ 363
+ ],
+ "score": 1.0,
+ "content": "prediction loss in the spectrogram domain, rather than the time domain, has the advantage of",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "increased invariance to phase differences between the generated and ground truth signals, which are",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 372,
+ 506,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 372,
+ 506,
+ 385
+ ],
+ "score": 1.0,
+ "content": "not perceptually salient. Seeing as the spectrogram extraction operation has several hyperparameters",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 384,
+ 504,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 384,
+ 504,
+ 395
+ ],
+ "score": 1.0,
+ "content": "and its implementation is not standardised, we provide the code we used for this in Appendix D. We",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 394,
+ 504,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 279,
+ 407
+ ],
+ "score": 1.0,
+ "content": "applied a small amount of jitter (by up to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 279,
+ 394,
+ 298,
+ 405
+ ],
+ "score": 0.86,
+ "content": "\\pm 6 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 299,
+ 394,
+ 345,
+ 407
+ ],
+ "score": 1.0,
+ "content": "samples at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 394,
+ 378,
+ 405
+ ],
+ "score": 0.72,
+ "content": "2 4 \\mathrm { k H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 378,
+ 394,
+ 504,
+ 407
+ ],
+ "score": 1.0,
+ "content": ") to the ground truth waveform",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 421,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 179,
+ 419
+ ],
+ "score": 1.0,
+ "content": "before computing",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 180,
+ 405,
+ 194,
+ 417
+ ],
+ "score": 0.9,
+ "content": "S _ { \\mathrm { g t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 195,
+ 405,
+ 421,
+ 419
+ ],
+ "score": 1.0,
+ "content": ", which helped to reduce artifacts in the generated audio.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 105,
+ 339,
+ 506,
+ 419
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 420,
+ 505,
+ 487
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 506,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 506,
+ 433
+ ],
+ "score": 1.0,
+ "content": "The inability to learn alignment from adversarial feedback alone is worth expanding on: likelihood-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 443
+ ],
+ "score": 1.0,
+ "content": "based autoregressive models have no issues learning alignment, because they are able to benefit",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 443,
+ 504,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 443,
+ 504,
+ 454
+ ],
+ "score": 1.0,
+ "content": "from teacher forcing (Williams & Zipser, 1989) during training: the model is trained to perform",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 466
+ ],
+ "score": 1.0,
+ "content": "next step prediction on each sequence step, given the preceding ground truth, and it is expected to",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 478
+ ],
+ "score": 1.0,
+ "content": "infer alignment only one step at a time. This is not compatible with feed-forward adversarial models",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 475,
+ 469,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 469,
+ 487
+ ],
+ "score": 1.0,
+ "content": "however, so the prediction loss is necessary to bootstrap alignment learning for our model.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 105,
+ 420,
+ 506,
+ 487
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 490,
+ 505,
+ 545
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 373,
+ 503
+ ],
+ "score": 1.0,
+ "content": "Note that although we make use of mel-spectrograms for training in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 374,
+ 490,
+ 398,
+ 502
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { L } _ { \\mathrm { p r e d } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 489,
+ 505,
+ 503
+ ],
+ "score": 1.0,
+ "content": "(and to compute the inputs",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 514
+ ],
+ "score": 1.0,
+ "content": "for the spectrogram discriminator, Section 2.3), the generator itself does not produce spectrograms",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 512,
+ 505,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 512,
+ 505,
+ 525
+ ],
+ "score": 1.0,
+ "content": "as part of the generation process. Rather, its outputs are raw waveforms, and we convert these",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 536
+ ],
+ "score": 1.0,
+ "content": "waveforms to spectrograms only for training (backpropagating gradients through the waveform to",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 535,
+ 267,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 535,
+ 267,
+ 547
+ ],
+ "score": 1.0,
+ "content": "mel-spectrogram conversion operation).",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 105,
+ 489,
+ 505,
+ 547
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 557,
+ 240,
+ 568
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 556,
+ 241,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 556,
+ 241,
+ 569
+ ],
+ "score": 1.0,
+ "content": "2.5 DYNAMIC TIME WARPING",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 575,
+ 505,
+ 642
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "The spectrogram prediction loss incorrectly assumes that token lengths are deterministic. We can",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "relax the requirement that the generated and ground truth spectrograms are exactly aligned, by",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "score": 1.0,
+ "content": "incorporating dynamic time warping (DTW) (Sakoe, 1971; Sakoe & Chiba, 1978). We calculate the",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 609,
+ 506,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 609,
+ 368,
+ 622
+ ],
+ "score": 1.0,
+ "content": "prediction loss by iteratively finding a minimal-cost alignment path",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 369,
+ 610,
+ 375,
+ 620
+ ],
+ "score": 0.82,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 609,
+ 506,
+ 622
+ ],
+ "score": 1.0,
+ "content": "between the generated and target",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 619,
+ 503,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 619,
+ 164,
+ 634
+ ],
+ "score": 1.0,
+ "content": "spectrograms,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 620,
+ 184,
+ 631
+ ],
+ "score": 0.9,
+ "content": "S _ { \\mathrm { g e n } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 619,
+ 202,
+ 634
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 203,
+ 620,
+ 217,
+ 632
+ ],
+ "score": 0.9,
+ "content": "S _ { \\mathrm { g t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 218,
+ 619,
+ 427,
+ 634
+ ],
+ "score": 1.0,
+ "content": ". We start at the first time step in both spectrograms:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 428,
+ 621,
+ 503,
+ 632
+ ],
+ "score": 0.92,
+ "content": "p _ { \\mathrm { g e n , 1 } } = p _ { \\mathrm { g t , 1 } } = 1",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 630,
+ 338,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 630,
+ 175,
+ 643
+ ],
+ "score": 1.0,
+ "content": "At each iteration",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 175,
+ 632,
+ 182,
+ 640
+ ],
+ "score": 0.81,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 630,
+ 338,
+ 643
+ ],
+ "score": 1.0,
+ "content": ", we take one of three possible actions:",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 105,
+ 576,
+ 506,
+ 643
+ ]
+ },
+ {
+ "type": "index",
+ "bbox": [
+ 130,
+ 649,
+ 493,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 647,
+ 491,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 647,
+ 266,
+ 664
+ ],
+ "score": 1.0,
+ "content": "1. go to the next time step in both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 266,
+ 649,
+ 491,
+ 662
+ ],
+ "score": 0.41,
+ "content": "S _ { \\mathrm { g e n } } , S _ { \\mathrm { g t } } \\colon p _ { \\mathrm { g e n } , k + 1 } = p _ { \\mathrm { g e n } , k } + 1 , p _ { \\mathrm { g t } , k + 1 } = p _ { \\mathrm { g t } , k } + 1 ;",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 44,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 128,
+ 663,
+ 452,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 663,
+ 248,
+ 678
+ ],
+ "score": 1.0,
+ "content": "2. go to the next time step in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 665,
+ 263,
+ 677
+ ],
+ "score": 0.89,
+ "content": "S _ { \\mathrm { g t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 663,
+ 286,
+ 678
+ ],
+ "score": 1.0,
+ "content": "only:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 287,
+ 665,
+ 448,
+ 677
+ ],
+ "score": 0.52,
+ "content": "p _ { \\mathrm { g e n } , k + 1 } = p _ { \\mathrm { g e n } , k } , p _ { \\mathrm { g t } , k + 1 } = p _ { \\mathrm { g t } , k } + 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 448,
+ 663,
+ 452,
+ 678
+ ],
+ "score": 1.0,
+ "content": ";",
+ "type": "text"
+ }
+ ],
+ "index": 45,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 128,
+ 678,
+ 456,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 678,
+ 248,
+ 694
+ ],
+ "score": 1.0,
+ "content": "3. go to the next time step in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 249,
+ 680,
+ 268,
+ 692
+ ],
+ "score": 0.91,
+ "content": "S _ { \\mathrm { g e n } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 268,
+ 678,
+ 293,
+ 694
+ ],
+ "score": 1.0,
+ "content": "only:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 293,
+ 681,
+ 383,
+ 692
+ ],
+ "score": 0.77,
+ "content": "p _ { \\mathrm { g e n } , k + 1 } = p _ { \\mathrm { g e n } , k } + 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 678,
+ 456,
+ 694
+ ],
+ "score": 1.0,
+ "content": ", pgt,k+1 = pgt,k.",
+ "type": "text"
+ }
+ ],
+ "index": 46,
+ "is_list_start_line": true
+ }
+ ],
+ "index": 45,
+ "bbox_fs": [
+ 128,
+ 647,
+ 491,
+ 694
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 698,
+ 503,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 196,
+ 713
+ ],
+ "score": 1.0,
+ "content": "The resulting path is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 699,
+ 372,
+ 712
+ ],
+ "score": 0.89,
+ "content": "p = \\langle ( p _ { \\mathrm { g e n } , 1 } , p _ { \\mathrm { g t } , 1 } ) , \\dots , ( p _ { \\mathrm { g e n } , K _ { p } } , p _ { \\mathrm { g t } , K _ { p } } ) \\rangle",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 372,
+ 697,
+ 405,
+ 713
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 699,
+ 420,
+ 711
+ ],
+ "score": 0.89,
+ "content": "K _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 420,
+ 697,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "is the length. Each",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 708,
+ 506,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 260,
+ 724
+ ],
+ "score": 1.0,
+ "content": "action is assigned a cost based on the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 260,
+ 711,
+ 272,
+ 721
+ ],
+ "score": 0.86,
+ "content": "L _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 708,
+ 346,
+ 724
+ ],
+ "score": 1.0,
+ "content": "distance between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 347,
+ 710,
+ 396,
+ 722
+ ],
+ "score": 0.91,
+ "content": "S _ { \\mathrm { g e n } } [ p _ { \\mathrm { g e n } , k } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 396,
+ 708,
+ 414,
+ 724
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 415,
+ 710,
+ 454,
+ 722
+ ],
+ "score": 0.92,
+ "content": "\\mathrm { \\dot { \\cal S } } _ { \\mathrm { g t } } [ p _ { \\mathrm { g t } , k } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 454,
+ 708,
+ 506,
+ 724
+ ],
+ "score": 1.0,
+ "content": ", and a warp",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 138,
+ 734
+ ],
+ "score": 1.0,
+ "content": "penalty",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 139,
+ 722,
+ 147,
+ 730
+ ],
+ "score": 0.71,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 147,
+ 720,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": "which is incurred if we choose not to advance both spectrograms in lockstep (i.e. we are",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 327,
+ 96
+ ],
+ "score": 1.0,
+ "content": "warping the spectrogram by taking action 2 or 3; we use",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 327,
+ 83,
+ 362,
+ 93
+ ],
+ "score": 0.88,
+ "content": "w = 1 . 0",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 363,
+ 82,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "). The warp penalty thus encourages",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 94,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 94,
+ 405,
+ 106
+ ],
+ "score": 1.0,
+ "content": "alignment paths that do not deviate too far from the identity alignment. Let",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 405,
+ 94,
+ 416,
+ 105
+ ],
+ "score": 0.87,
+ "content": "\\delta _ { k }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 416,
+ 94,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "be an indicator which",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 104,
+ 458,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 104,
+ 434,
+ 118
+ ],
+ "score": 1.0,
+ "content": "is 1 for iterations where warping occurs, and 0 otherwise. Then the total path cost",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 434,
+ 106,
+ 444,
+ 117
+ ],
+ "score": 0.87,
+ "content": "c _ { p }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 445,
+ 104,
+ 458,
+ 118
+ ],
+ "score": 1.0,
+ "content": "is:",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 48,
+ "bbox_fs": [
+ 105,
+ 697,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 327,
+ 96
+ ],
+ "score": 1.0,
+ "content": "warping the spectrogram by taking action 2 or 3; we use",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 83,
+ 362,
+ 93
+ ],
+ "score": 0.88,
+ "content": "w = 1 . 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 363,
+ 82,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "). The warp penalty thus encourages",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 94,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 94,
+ 405,
+ 106
+ ],
+ "score": 1.0,
+ "content": "alignment paths that do not deviate too far from the identity alignment. Let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 405,
+ 94,
+ 416,
+ 105
+ ],
+ "score": 0.87,
+ "content": "\\delta _ { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 416,
+ 94,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "be an indicator which",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 104,
+ 458,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 104,
+ 434,
+ 118
+ ],
+ "score": 1.0,
+ "content": "is 1 for iterations where warping occurs, and 0 otherwise. Then the total path cost",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 434,
+ 106,
+ 444,
+ 117
+ ],
+ "score": 0.87,
+ "content": "c _ { p }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 445,
+ 104,
+ 458,
+ 118
+ ],
+ "score": 1.0,
+ "content": "is:",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 172,
+ 118,
+ 439,
+ 140
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 118,
+ 439,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 118,
+ 439,
+ 140
+ ],
+ "score": 0.91,
+ "content": "\\begin{array} { r } { c _ { p } = \\sum _ { k = 1 } ^ { K _ { p } } \\Big ( \\boldsymbol { w } \\cdot \\boldsymbol { \\delta _ { k } } + \\frac { 1 } { F } \\sum _ { f = 1 } ^ { F } | S _ { \\mathrm { g e n } } [ p _ { \\mathrm { g e n } , k } , f ] - S _ { \\mathrm { g t } } [ p _ { \\mathrm { g t } , k } , f ] | \\Big ) . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "5e55027ea6418ddbbb991219c835af82b42c74a3172fd85b36bb0223dc35e50a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 172,
+ 118,
+ 439,
+ 140
+ ],
+ "spans": [],
+ "index": 3
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 118,
+ 141,
+ 478,
+ 154
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 140,
+ 481,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 140,
+ 260,
+ 156
+ ],
+ "score": 1.0,
+ "content": "depends on the degree of warping",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 260,
+ 142,
+ 343,
+ 154
+ ],
+ "score": 0.93,
+ "content": "( T \\leq K _ { p } \\leq 2 T - 1 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 140,
+ 481,
+ 156
+ ],
+ "score": 1.0,
+ "content": ". The DTW prediction loss is then:",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 271,
+ 156,
+ 339,
+ 174
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 271,
+ 156,
+ 339,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 271,
+ 156,
+ 339,
+ 174
+ ],
+ "score": 0.9,
+ "content": "\\mathcal { L } _ { \\mathrm { p r e d } } ^ { \\prime } = \\operatorname* { m i n } _ { p \\in \\mathcal { P } } c _ { p } ,",
+ "type": "interline_equation",
+ "image_path": "a0274c1bfcce379ed130b205b04215c357f6e780ecd4665614c4bcf802543ff3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 271,
+ 156,
+ 339,
+ 174
+ ],
+ "spans": [],
+ "index": 5
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 210
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 175,
+ 502,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 175,
+ 132,
+ 192
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 177,
+ 141,
+ 187
+ ],
+ "score": 0.84,
+ "content": "\\mathcal { P }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 141,
+ 175,
+ 248,
+ 192
+ ],
+ "score": 1.0,
+ "content": "is the set of all valid paths.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 177,
+ 274,
+ 188
+ ],
+ "score": 0.91,
+ "content": "p \\in \\mathcal P",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 175,
+ 318,
+ 192
+ ],
+ "score": 1.0,
+ "content": "only when",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 318,
+ 178,
+ 394,
+ 189
+ ],
+ "score": 0.91,
+ "content": "p _ { \\mathrm { g e n , 1 } } = p _ { \\mathrm { g t , 1 } } = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 394,
+ 175,
+ 411,
+ 192
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 411,
+ 177,
+ 502,
+ 190
+ ],
+ "score": 0.9,
+ "content": "p _ { \\mathrm { g e n } , K _ { p } } = p _ { \\mathrm { g t } , K _ { p } } = T",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 188,
+ 504,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 188,
+ 504,
+ 200
+ ],
+ "score": 1.0,
+ "content": "i.e. the first and last timesteps of the spectrograms are aligned. To find the minimum, we use dynamic",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 488,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 488,
+ 212
+ ],
+ "score": 1.0,
+ "content": "programming. Figure 2 shows a diagram of an optimal alignment path between two sequences.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 214,
+ 376,
+ 258
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 212,
+ 378,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 212,
+ 378,
+ 227
+ ],
+ "score": 1.0,
+ "content": "DTW is differentiable, but the minimum across all paths makes op-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 225,
+ 376,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 376,
+ 237
+ ],
+ "score": 1.0,
+ "content": "timisation difficult, because the gradient is propagated only through",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 236,
+ 377,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 236,
+ 377,
+ 247
+ ],
+ "score": 1.0,
+ "content": "the minimal path. We use a soft version of DTW instead (Cuturi &",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 246,
+ 377,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 246,
+ 377,
+ 258
+ ],
+ "score": 1.0,
+ "content": "Blondel, 2017), which replaces the minimum with the soft minimum:",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 166,
+ 260,
+ 315,
+ 276
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 166,
+ 260,
+ 315,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 166,
+ 260,
+ 315,
+ 276
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { p r e d } } ^ { \\prime \\prime } = - \\tau \\cdot \\log \\sum _ { p \\in \\mathcal { P } } \\exp \\left( - \\frac { c _ { p } } { \\tau } \\right) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "63b0bbdd0ea571b873711903c5b179bb9f83e424aee2ca04f4637076877dcb55.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 166,
+ 260,
+ 315,
+ 276
+ ],
+ "spans": [],
+ "index": 13
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 277,
+ 376,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 277,
+ 377,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 277,
+ 132,
+ 290
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 133,
+ 278,
+ 171,
+ 288
+ ],
+ "score": 0.9,
+ "content": "\\tau = 0 . 0 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 277,
+ 377,
+ 290
+ ],
+ "score": 1.0,
+ "content": "is a temperature parameter and the loss scale factor",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 107,
+ 288,
+ 376,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 289,
+ 159,
+ 300
+ ],
+ "score": 0.88,
+ "content": "\\lambda _ { \\mathrm { p r e d } } = 1 . 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 160,
+ 288,
+ 376,
+ 300
+ ],
+ "score": 1.0,
+ "content": ". Note that the minimum operation is recovered by",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 299,
+ 377,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 299,
+ 135,
+ 312
+ ],
+ "score": 1.0,
+ "content": "letting",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 136,
+ 300,
+ 166,
+ 310
+ ],
+ "score": 0.89,
+ "content": "\\tau 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 167,
+ 299,
+ 377,
+ 312
+ ],
+ "score": 1.0,
+ "content": ". The resulting loss is a weighted aggregated cost",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 310,
+ 376,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 376,
+ 322
+ ],
+ "score": 1.0,
+ "content": "across all paths, enabling gradient propagation through all feasible",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 321,
+ 378,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 263,
+ 334
+ ],
+ "score": 1.0,
+ "content": "paths. This creates a trade-off: a higher",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 263,
+ 324,
+ 270,
+ 332
+ ],
+ "score": 0.76,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 271,
+ 321,
+ 378,
+ 334
+ ],
+ "score": 1.0,
+ "content": "makes optimisation easier,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 332,
+ 378,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 378,
+ 344
+ ],
+ "score": 1.0,
+ "content": "but the resulting loss less accurately reflects the minimal path cost.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 343,
+ 378,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 378,
+ 356
+ ],
+ "score": 1.0,
+ "content": "Pseudocode for the soft DTW procedure is provided in Appendix E.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 358,
+ 378,
+ 425
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 378,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 378,
+ 371
+ ],
+ "score": 1.0,
+ "content": "By relaxing alignment in the prediction loss, the generator can pro-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 368,
+ 377,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 368,
+ 377,
+ 382
+ ],
+ "score": 1.0,
+ "content": "duce waveforms that are not exactly aligned, without being heavily",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 380,
+ 378,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 378,
+ 393
+ ],
+ "score": 1.0,
+ "content": "penalised for it. This creates a synergy with the adversarial loss:",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 376,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 376,
+ 404
+ ],
+ "score": 1.0,
+ "content": "instead of working against each other because of the rigidity of the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 402,
+ 376,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 376,
+ 414
+ ],
+ "score": 1.0,
+ "content": "prediction loss, the losses now cooperate to reward realistic audio",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 413,
+ 376,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 376,
+ 426
+ ],
+ "score": 1.0,
+ "content": "generation with stochastic alignment. Note that the prediction loss",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 386,
+ 227,
+ 504,
+ 346
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 386,
+ 227,
+ 504,
+ 346
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 386,
+ 227,
+ 504,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 386,
+ 227,
+ 504,
+ 346
+ ],
+ "score": 0.961,
+ "type": "image",
+ "image_path": "3ec946c31ee94122d585bbdb325279673d19931189642e35e802f93e8666c4d4.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 27.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 386,
+ 227,
+ 504,
+ 286.5
+ ],
+ "spans": [],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 386,
+ 286.5,
+ 504,
+ 346.0
+ ],
+ "spans": [],
+ "index": 28
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 383,
+ 352,
+ 505,
+ 417
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 383,
+ 351,
+ 506,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 351,
+ 506,
+ 364
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Dynamic time warp-",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 382,
+ 362,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 382,
+ 362,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "ing between two sequences",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 383,
+ 373,
+ 506,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 373,
+ 506,
+ 386
+ ],
+ "score": 1.0,
+ "content": "finds a minimal-cost align-",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 382,
+ 384,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 382,
+ 384,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "ment path. Positions where",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 383,
+ 396,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 396,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "warping occurs are marked",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 383,
+ 406,
+ 441,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 406,
+ 441,
+ 418
+ ],
+ "score": 1.0,
+ "content": "with a border.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 31.5
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 425,
+ 501,
+ 457
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 503,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 503,
+ 436
+ ],
+ "score": 1.0,
+ "content": "is computed on a training window, and not on full length utterances, so we still assume that the start",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 435,
+ 503,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 503,
+ 447
+ ],
+ "score": 1.0,
+ "content": "and end points of the windows are exactly aligned. While this might be incorrect, it does not seem to",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 446,
+ 243,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 243,
+ 458
+ ],
+ "score": 1.0,
+ "content": "be much of a problem in practice.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 468,
+ 233,
+ 479
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 468,
+ 234,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 468,
+ 234,
+ 481
+ ],
+ "score": 1.0,
+ "content": "2.6 ALIGNER LENGTH LOSS",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 487,
+ 505,
+ 532
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 505,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 505,
+ 500
+ ],
+ "score": 1.0,
+ "content": "To ensure that the model produces realistic token length predictions, we add a loss which encourages",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 497,
+ 506,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 506,
+ 512
+ ],
+ "score": 1.0,
+ "content": "the predicted utterance length to be close to the ground truth length. This length is found by summing",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 509,
+ 506,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 510,
+ 234,
+ 521
+ ],
+ "score": 1.0,
+ "content": "all token length predictions. Let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 510,
+ 242,
+ 519
+ ],
+ "score": 0.82,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 242,
+ 510,
+ 473,
+ 521
+ ],
+ "score": 1.0,
+ "content": "be the the number of time steps in the training utterance at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 509,
+ 502,
+ 520
+ ],
+ "score": 0.7,
+ "content": "2 0 0 \\mathrm { H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 503,
+ 510,
+ 506,
+ 521
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 107,
+ 520,
+ 478,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 521,
+ 116,
+ 531
+ ],
+ "score": 0.87,
+ "content": "l _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 116,
+ 520,
+ 224,
+ 533
+ ],
+ "score": 1.0,
+ "content": "the predicted length of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 522,
+ 231,
+ 530
+ ],
+ "score": 0.66,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 231,
+ 520,
+ 284,
+ 533
+ ],
+ "score": 1.0,
+ "content": "th token, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 285,
+ 520,
+ 295,
+ 530
+ ],
+ "score": 0.82,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 295,
+ 520,
+ 478,
+ 533
+ ],
+ "score": 1.0,
+ "content": "the number of tokens, then the length loss is:",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 241,
+ 533,
+ 370,
+ 557
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 241,
+ 533,
+ 370,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 241,
+ 533,
+ 370,
+ 557
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { l e n g t h } } = \\frac { 1 } { 2 } \\left( L - \\sum _ { n = 1 } ^ { N } l _ { n } \\right) ^ { 2 } . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "cfe00b01bfe2a11b352c8e3753171589565351b9c39d23dd72a5cfe5844cb4d9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 43,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 241,
+ 533,
+ 370,
+ 557
+ ],
+ "spans": [],
+ "index": 43
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 559,
+ 505,
+ 581
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 559,
+ 505,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 559,
+ 190,
+ 571
+ ],
+ "score": 1.0,
+ "content": "We use a scale factor",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 190,
+ 559,
+ 245,
+ 570
+ ],
+ "score": 0.9,
+ "content": "\\lambda _ { \\mathrm { l e n g t h } } = 0 . 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 559,
+ 439,
+ 571
+ ],
+ "score": 1.0,
+ "content": ". Note that we cannot match the predicted lengths",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 440,
+ 559,
+ 449,
+ 570
+ ],
+ "score": 0.88,
+ "content": "l _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 559,
+ 505,
+ 571
+ ],
+ "score": 1.0,
+ "content": "to the ground",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 569,
+ 354,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 569,
+ 354,
+ 582
+ ],
+ "score": 1.0,
+ "content": "truth lengths individually, because the latter are not available.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 44.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 592,
+ 232,
+ 603
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 592,
+ 231,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 592,
+ 231,
+ 604
+ ],
+ "score": 1.0,
+ "content": "2.7 TEXT PRE-PROCESSING",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 46
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 506,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 506,
+ 624
+ ],
+ "score": 1.0,
+ "content": "Although our model works well with character input, we find that sample quality improves signifi-",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": "cantly using phoneme input instead. This is not too surprising, given the heterogeneous way in which",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 646
+ ],
+ "score": 1.0,
+ "content": "spellings map to phonemes, particularly in the English language. Many character sequences also",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": "have special pronunciations, such as numbers, dates, units of measurement and website domains, and",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 104,
+ 654,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 654,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "a very large training dataset would be required for the model to learn to pronounce these correctly.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "Text normalisation (Zhang et al., 2019) can be applied beforehand to spell out these sequences as",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "they are typically pronounced (e.g., 1976 could become nineteen seventy six), potentially followed by",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 701
+ ],
+ "score": 1.0,
+ "content": "conversion to phonemes. We use an open source tool, phonemizer (Bernard, 2020), which performs",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "partial normalisation and phonemisation (see Appendix F). Finally, whether we train on text or",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "phoneme input sequences, we pre- and post-pad the sequence with a special silence token (for training",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 507,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 507,
+ 734
+ ],
+ "score": 1.0,
+ "content": "and inference), to allow the aligner to account for silence at the beginning and end of each utterance.",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 26,
+ 292,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 116
+ ],
+ "lines": [],
+ "index": 1,
+ "bbox_fs": [
+ 104,
+ 82,
+ 505,
+ 118
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 172,
+ 118,
+ 439,
+ 140
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 118,
+ 439,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 118,
+ 439,
+ 140
+ ],
+ "score": 0.91,
+ "content": "\\begin{array} { r } { c _ { p } = \\sum _ { k = 1 } ^ { K _ { p } } \\Big ( \\boldsymbol { w } \\cdot \\boldsymbol { \\delta _ { k } } + \\frac { 1 } { F } \\sum _ { f = 1 } ^ { F } | S _ { \\mathrm { g e n } } [ p _ { \\mathrm { g e n } , k } , f ] - S _ { \\mathrm { g t } } [ p _ { \\mathrm { g t } , k } , f ] | \\Big ) . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "5e55027ea6418ddbbb991219c835af82b42c74a3172fd85b36bb0223dc35e50a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 172,
+ 118,
+ 439,
+ 140
+ ],
+ "spans": [],
+ "index": 3
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 118,
+ 141,
+ 478,
+ 154
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 140,
+ 481,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 140,
+ 260,
+ 156
+ ],
+ "score": 1.0,
+ "content": "depends on the degree of warping",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 260,
+ 142,
+ 343,
+ 154
+ ],
+ "score": 0.93,
+ "content": "( T \\leq K _ { p } \\leq 2 T - 1 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 140,
+ 481,
+ 156
+ ],
+ "score": 1.0,
+ "content": ". The DTW prediction loss is then:",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 121,
+ 140,
+ 481,
+ 156
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 271,
+ 156,
+ 339,
+ 174
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 271,
+ 156,
+ 339,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 271,
+ 156,
+ 339,
+ 174
+ ],
+ "score": 0.9,
+ "content": "\\mathcal { L } _ { \\mathrm { p r e d } } ^ { \\prime } = \\operatorname* { m i n } _ { p \\in \\mathcal { P } } c _ { p } ,",
+ "type": "interline_equation",
+ "image_path": "a0274c1bfcce379ed130b205b04215c357f6e780ecd4665614c4bcf802543ff3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 271,
+ 156,
+ 339,
+ 174
+ ],
+ "spans": [],
+ "index": 5
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 210
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 175,
+ 502,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 175,
+ 132,
+ 192
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 177,
+ 141,
+ 187
+ ],
+ "score": 0.84,
+ "content": "\\mathcal { P }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 141,
+ 175,
+ 248,
+ 192
+ ],
+ "score": 1.0,
+ "content": "is the set of all valid paths.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 177,
+ 274,
+ 188
+ ],
+ "score": 0.91,
+ "content": "p \\in \\mathcal P",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 175,
+ 318,
+ 192
+ ],
+ "score": 1.0,
+ "content": "only when",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 318,
+ 178,
+ 394,
+ 189
+ ],
+ "score": 0.91,
+ "content": "p _ { \\mathrm { g e n , 1 } } = p _ { \\mathrm { g t , 1 } } = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 394,
+ 175,
+ 411,
+ 192
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 411,
+ 177,
+ 502,
+ 190
+ ],
+ "score": 0.9,
+ "content": "p _ { \\mathrm { g e n } , K _ { p } } = p _ { \\mathrm { g t } , K _ { p } } = T",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 188,
+ 504,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 188,
+ 504,
+ 200
+ ],
+ "score": 1.0,
+ "content": "i.e. the first and last timesteps of the spectrograms are aligned. To find the minimum, we use dynamic",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 488,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 488,
+ 212
+ ],
+ "score": 1.0,
+ "content": "programming. Figure 2 shows a diagram of an optimal alignment path between two sequences.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7,
+ "bbox_fs": [
+ 104,
+ 175,
+ 504,
+ 212
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 214,
+ 376,
+ 258
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 212,
+ 378,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 212,
+ 378,
+ 227
+ ],
+ "score": 1.0,
+ "content": "DTW is differentiable, but the minimum across all paths makes op-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 225,
+ 376,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 376,
+ 237
+ ],
+ "score": 1.0,
+ "content": "timisation difficult, because the gradient is propagated only through",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 236,
+ 377,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 236,
+ 377,
+ 247
+ ],
+ "score": 1.0,
+ "content": "the minimal path. We use a soft version of DTW instead (Cuturi &",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 246,
+ 377,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 246,
+ 377,
+ 258
+ ],
+ "score": 1.0,
+ "content": "Blondel, 2017), which replaces the minimum with the soft minimum:",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10.5,
+ "bbox_fs": [
+ 105,
+ 212,
+ 378,
+ 258
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 166,
+ 260,
+ 315,
+ 276
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 166,
+ 260,
+ 315,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 166,
+ 260,
+ 315,
+ 276
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { p r e d } } ^ { \\prime \\prime } = - \\tau \\cdot \\log \\sum _ { p \\in \\mathcal { P } } \\exp \\left( - \\frac { c _ { p } } { \\tau } \\right) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "63b0bbdd0ea571b873711903c5b179bb9f83e424aee2ca04f4637076877dcb55.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 166,
+ 260,
+ 315,
+ 276
+ ],
+ "spans": [],
+ "index": 13
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 277,
+ 376,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 277,
+ 377,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 277,
+ 132,
+ 290
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 133,
+ 278,
+ 171,
+ 288
+ ],
+ "score": 0.9,
+ "content": "\\tau = 0 . 0 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 277,
+ 377,
+ 290
+ ],
+ "score": 1.0,
+ "content": "is a temperature parameter and the loss scale factor",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 107,
+ 288,
+ 376,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 289,
+ 159,
+ 300
+ ],
+ "score": 0.88,
+ "content": "\\lambda _ { \\mathrm { p r e d } } = 1 . 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 160,
+ 288,
+ 376,
+ 300
+ ],
+ "score": 1.0,
+ "content": ". Note that the minimum operation is recovered by",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 299,
+ 377,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 299,
+ 135,
+ 312
+ ],
+ "score": 1.0,
+ "content": "letting",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 136,
+ 300,
+ 166,
+ 310
+ ],
+ "score": 0.89,
+ "content": "\\tau 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 167,
+ 299,
+ 377,
+ 312
+ ],
+ "score": 1.0,
+ "content": ". The resulting loss is a weighted aggregated cost",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 310,
+ 376,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 376,
+ 322
+ ],
+ "score": 1.0,
+ "content": "across all paths, enabling gradient propagation through all feasible",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 321,
+ 378,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 263,
+ 334
+ ],
+ "score": 1.0,
+ "content": "paths. This creates a trade-off: a higher",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 263,
+ 324,
+ 270,
+ 332
+ ],
+ "score": 0.76,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 271,
+ 321,
+ 378,
+ 334
+ ],
+ "score": 1.0,
+ "content": "makes optimisation easier,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 332,
+ 378,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 378,
+ 344
+ ],
+ "score": 1.0,
+ "content": "but the resulting loss less accurately reflects the minimal path cost.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 343,
+ 378,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 378,
+ 356
+ ],
+ "score": 1.0,
+ "content": "Pseudocode for the soft DTW procedure is provided in Appendix E.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 17,
+ "bbox_fs": [
+ 105,
+ 277,
+ 378,
+ 356
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 358,
+ 378,
+ 425
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 378,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 378,
+ 371
+ ],
+ "score": 1.0,
+ "content": "By relaxing alignment in the prediction loss, the generator can pro-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 368,
+ 377,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 368,
+ 377,
+ 382
+ ],
+ "score": 1.0,
+ "content": "duce waveforms that are not exactly aligned, without being heavily",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 380,
+ 378,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 378,
+ 393
+ ],
+ "score": 1.0,
+ "content": "penalised for it. This creates a synergy with the adversarial loss:",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 376,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 376,
+ 404
+ ],
+ "score": 1.0,
+ "content": "instead of working against each other because of the rigidity of the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 402,
+ 376,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 376,
+ 414
+ ],
+ "score": 1.0,
+ "content": "prediction loss, the losses now cooperate to reward realistic audio",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 413,
+ 376,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 413,
+ 376,
+ 426
+ ],
+ "score": 1.0,
+ "content": "generation with stochastic alignment. Note that the prediction loss",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 503,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 503,
+ 436
+ ],
+ "score": 1.0,
+ "content": "is computed on a training window, and not on full length utterances, so we still assume that the start",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 435,
+ 503,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 503,
+ 447
+ ],
+ "score": 1.0,
+ "content": "and end points of the windows are exactly aligned. While this might be incorrect, it does not seem to",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 446,
+ 243,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 243,
+ 458
+ ],
+ "score": 1.0,
+ "content": "be much of a problem in practice.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 105,
+ 358,
+ 378,
+ 426
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 386,
+ 227,
+ 504,
+ 346
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 386,
+ 227,
+ 504,
+ 346
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 386,
+ 227,
+ 504,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 386,
+ 227,
+ 504,
+ 346
+ ],
+ "score": 0.961,
+ "type": "image",
+ "image_path": "3ec946c31ee94122d585bbdb325279673d19931189642e35e802f93e8666c4d4.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 27.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 386,
+ 227,
+ 504,
+ 286.5
+ ],
+ "spans": [],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 386,
+ 286.5,
+ 504,
+ 346.0
+ ],
+ "spans": [],
+ "index": 28
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 383,
+ 352,
+ 505,
+ 417
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 383,
+ 351,
+ 506,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 351,
+ 506,
+ 364
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Dynamic time warp-",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 382,
+ 362,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 382,
+ 362,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "ing between two sequences",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 383,
+ 373,
+ 506,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 373,
+ 506,
+ 386
+ ],
+ "score": 1.0,
+ "content": "finds a minimal-cost align-",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 382,
+ 384,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 382,
+ 384,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "ment path. Positions where",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 383,
+ 396,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 396,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "warping occurs are marked",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 383,
+ 406,
+ 441,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 406,
+ 441,
+ 418
+ ],
+ "score": 1.0,
+ "content": "with a border.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 31.5
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 425,
+ 501,
+ 457
+ ],
+ "lines": [],
+ "index": 36,
+ "bbox_fs": [
+ 105,
+ 424,
+ 503,
+ 458
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 468,
+ 233,
+ 479
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 468,
+ 234,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 468,
+ 234,
+ 481
+ ],
+ "score": 1.0,
+ "content": "2.6 ALIGNER LENGTH LOSS",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 487,
+ 505,
+ 532
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 505,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 486,
+ 505,
+ 500
+ ],
+ "score": 1.0,
+ "content": "To ensure that the model produces realistic token length predictions, we add a loss which encourages",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 497,
+ 506,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 506,
+ 512
+ ],
+ "score": 1.0,
+ "content": "the predicted utterance length to be close to the ground truth length. This length is found by summing",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 509,
+ 506,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 510,
+ 234,
+ 521
+ ],
+ "score": 1.0,
+ "content": "all token length predictions. Let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 510,
+ 242,
+ 519
+ ],
+ "score": 0.82,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 242,
+ 510,
+ 473,
+ 521
+ ],
+ "score": 1.0,
+ "content": "be the the number of time steps in the training utterance at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 509,
+ 502,
+ 520
+ ],
+ "score": 0.7,
+ "content": "2 0 0 \\mathrm { H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 503,
+ 510,
+ 506,
+ 521
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 107,
+ 520,
+ 478,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 521,
+ 116,
+ 531
+ ],
+ "score": 0.87,
+ "content": "l _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 116,
+ 520,
+ 224,
+ 533
+ ],
+ "score": 1.0,
+ "content": "the predicted length of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 522,
+ 231,
+ 530
+ ],
+ "score": 0.66,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 231,
+ 520,
+ 284,
+ 533
+ ],
+ "score": 1.0,
+ "content": "th token, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 285,
+ 520,
+ 295,
+ 530
+ ],
+ "score": 0.82,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 295,
+ 520,
+ 478,
+ 533
+ ],
+ "score": 1.0,
+ "content": "the number of tokens, then the length loss is:",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 105,
+ 486,
+ 506,
+ 533
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 241,
+ 533,
+ 370,
+ 557
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 241,
+ 533,
+ 370,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 241,
+ 533,
+ 370,
+ 557
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { l e n g t h } } = \\frac { 1 } { 2 } \\left( L - \\sum _ { n = 1 } ^ { N } l _ { n } \\right) ^ { 2 } . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "cfe00b01bfe2a11b352c8e3753171589565351b9c39d23dd72a5cfe5844cb4d9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 43,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 241,
+ 533,
+ 370,
+ 557
+ ],
+ "spans": [],
+ "index": 43
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 559,
+ 505,
+ 581
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 559,
+ 505,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 559,
+ 190,
+ 571
+ ],
+ "score": 1.0,
+ "content": "We use a scale factor",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 190,
+ 559,
+ 245,
+ 570
+ ],
+ "score": 0.9,
+ "content": "\\lambda _ { \\mathrm { l e n g t h } } = 0 . 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 559,
+ 439,
+ 571
+ ],
+ "score": 1.0,
+ "content": ". Note that we cannot match the predicted lengths",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 440,
+ 559,
+ 449,
+ 570
+ ],
+ "score": 0.88,
+ "content": "l _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 559,
+ 505,
+ 571
+ ],
+ "score": 1.0,
+ "content": "to the ground",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 569,
+ 354,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 569,
+ 354,
+ 582
+ ],
+ "score": 1.0,
+ "content": "truth lengths individually, because the latter are not available.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 44.5,
+ "bbox_fs": [
+ 106,
+ 559,
+ 505,
+ 582
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 592,
+ 232,
+ 603
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 592,
+ 231,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 592,
+ 231,
+ 604
+ ],
+ "score": 1.0,
+ "content": "2.7 TEXT PRE-PROCESSING",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 46
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 506,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 506,
+ 624
+ ],
+ "score": 1.0,
+ "content": "Although our model works well with character input, we find that sample quality improves signifi-",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": "cantly using phoneme input instead. This is not too surprising, given the heterogeneous way in which",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 646
+ ],
+ "score": 1.0,
+ "content": "spellings map to phonemes, particularly in the English language. Many character sequences also",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": "have special pronunciations, such as numbers, dates, units of measurement and website domains, and",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 104,
+ 654,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 654,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "a very large training dataset would be required for the model to learn to pronounce these correctly.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "Text normalisation (Zhang et al., 2019) can be applied beforehand to spell out these sequences as",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "they are typically pronounced (e.g., 1976 could become nineteen seventy six), potentially followed by",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 701
+ ],
+ "score": 1.0,
+ "content": "conversion to phonemes. We use an open source tool, phonemizer (Bernard, 2020), which performs",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "partial normalisation and phonemisation (see Appendix F). Finally, whether we train on text or",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "phoneme input sequences, we pre- and post-pad the sequence with a special silence token (for training",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 507,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 507,
+ 734
+ ],
+ "score": 1.0,
+ "content": "and inference), to allow the aligner to account for silence at the beginning and end of each utterance.",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ }
+ ],
+ "index": 52,
+ "bbox_fs": [
+ 104,
+ 610,
+ 507,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 81,
+ 208,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 79,
+ 211,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 79,
+ 211,
+ 96
+ ],
+ "score": 1.0,
+ "content": "3 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 106,
+ 505,
+ 216
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 120
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 120
+ ],
+ "score": 1.0,
+ "content": "Speech generation saw significant quality improvements once treating it as a generative modelling",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 505,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 505,
+ 130
+ ],
+ "score": 1.0,
+ "content": "problem became the norm (Zen et al., 2009; van den Oord et al., 2016). Likelihood-based approaches",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 142
+ ],
+ "score": 1.0,
+ "content": "dominate, but generative adversarial networks (GANs) (Goodfellow et al., 2014) have been making",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 138,
+ 506,
+ 152
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 506,
+ 152
+ ],
+ "score": 1.0,
+ "content": "significant inroads recently. A common thread through most of the literature is a separation of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 150,
+ 506,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 150,
+ 506,
+ 163
+ ],
+ "score": 1.0,
+ "content": "the speech generation process into multiple stages: coarse-grained temporally aligned intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 161,
+ 506,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 161,
+ 506,
+ 174
+ ],
+ "score": 1.0,
+ "content": "representations, such as mel-spectrograms, are used to divide the task into more manageable sub-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 171,
+ 506,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 506,
+ 185
+ ],
+ "score": 1.0,
+ "content": "problems. Many works focus exclusively on either spectrogram generation or vocoding (generating a",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 183,
+ 506,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 183,
+ 506,
+ 195
+ ],
+ "score": 1.0,
+ "content": "waveform from a spectrogram). Our work is different in this respect, and we will point out which",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 194,
+ 505,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 194,
+ 505,
+ 207
+ ],
+ "score": 1.0,
+ "content": "stages of the generation process are addressed by each model. In Appendix J, Table 6 we compare",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 204,
+ 428,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 428,
+ 217
+ ],
+ "score": 1.0,
+ "content": "these methods in terms of the inputs and outputs to each stage of their pipelines.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 219,
+ 506,
+ 340
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 233
+ ],
+ "score": 1.0,
+ "content": "Initially, most likelihood-based models for TTS were autoregressive (van den Oord et al., 2016; Mehri",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 231,
+ 506,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 231,
+ 506,
+ 243
+ ],
+ "score": 1.0,
+ "content": "et al., 2017; Arik et al., 2017), which means that there is a sequential dependency between subsequent",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "time steps of the produced output signal. That makes these models impractical for real-time use,",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 507,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 507,
+ 266
+ ],
+ "score": 1.0,
+ "content": "although this can be addressed with careful engineering (Kalchbrenner et al., 2018; Valin & Skoglund,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 263,
+ 506,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 506,
+ 276
+ ],
+ "score": 1.0,
+ "content": "2019). More recently, flow-based models (Papamakarios et al., 2019) have been explored as a",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 274,
+ 506,
+ 287
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 506,
+ 287
+ ],
+ "score": 1.0,
+ "content": "feed-forward alternative that enables fast inference (without sequential dependencies). These can",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 285,
+ 506,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 285,
+ 506,
+ 298
+ ],
+ "score": 1.0,
+ "content": "either be trained directly using maximum likelihood (Prenger et al., 2019; Kim et al., 2019; Ping et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 295,
+ 507,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 295,
+ 507,
+ 310
+ ],
+ "score": 1.0,
+ "content": "2019b), or through distillation from an autoregressive model (van den Oord et al., 2018; Ping et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 307,
+ 506,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 307,
+ 506,
+ 320
+ ],
+ "score": 1.0,
+ "content": "2019a). All of these models produce waveforms conditioned on an intermediate representation: either",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 318,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "spectrograms or “linguistic features”, which contain temporally-aligned high-level information about",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 329,
+ 491,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 329,
+ 491,
+ 342
+ ],
+ "score": 1.0,
+ "content": "the speech signal. Spectrogram-conditioned waveform models are often referred to as vocoders.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 344,
+ 505,
+ 432
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 506,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 506,
+ 357
+ ],
+ "score": 1.0,
+ "content": "A growing body of work has applied GAN (Goodfellow et al., 2014) variants to speech synthe-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 355,
+ 506,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 355,
+ 506,
+ 367
+ ],
+ "score": 1.0,
+ "content": "sis (Donahue et al., 2019). An important advantage of adversarial losses for TTS is a focus on",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 366,
+ 506,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 506,
+ 378
+ ],
+ "score": 1.0,
+ "content": "realism over diversity; the latter is less important in this setting. This enables a more efficient use of",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 378,
+ 506,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 506,
+ 390
+ ],
+ "score": 1.0,
+ "content": "capacity compared to models trained with maximum likelihood. MelGAN (Kumar et al., 2019) and",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 387,
+ 505,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 505,
+ 401
+ ],
+ "score": 1.0,
+ "content": "Parallel WaveGAN (Yamamoto et al., 2020) are adversarial vocoders, producing raw waveforms from",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 398,
+ 507,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 398,
+ 507,
+ 413
+ ],
+ "score": 1.0,
+ "content": "mel-spectrograms. Neekhara et al. (2019) predict magnitude spectrograms from mel-spectrograms.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 410,
+ 505,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 505,
+ 422
+ ],
+ "score": 1.0,
+ "content": "Most directly related to our work is GAN-TTS (Binkowski et al., 2020), which produces waveforms ´",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 421,
+ 395,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 421,
+ 395,
+ 433
+ ],
+ "score": 1.0,
+ "content": "conditioned on aligned linguistic features, and we build upon that work.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 436,
+ 505,
+ 601
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 436,
+ 506,
+ 449
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 436,
+ 506,
+ 449
+ ],
+ "score": 1.0,
+ "content": "Another important line of work covers spectrogram generation from text. Such models rely on a",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 447,
+ 506,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 447,
+ 506,
+ 460
+ ],
+ "score": 1.0,
+ "content": "vocoder to convert the spectrograms into waveforms (for which one of the previously mentioned",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 458,
+ 506,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 458,
+ 506,
+ 471
+ ],
+ "score": 1.0,
+ "content": "models could be used, or a traditional spectrogram inversion technique (Griffin & Lim, 1984)).",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 467,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 467,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "Tacotron 1 & 2 (Wang et al., 2017; Shen et al., 2018), Deep Voice 2 & 3 (Gibiansky et al., 2017; Ping",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 104,
+ 479,
+ 506,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 479,
+ 506,
+ 494
+ ],
+ "score": 1.0,
+ "content": "et al., 2018), TransformerTTS (Li et al., 2019), Flowtron (Valle et al., 2020), and VoiceLoop (Taigman",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 491,
+ 507,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 491,
+ 507,
+ 504
+ ],
+ "score": 1.0,
+ "content": "et al., 2017) are autoregressive models that generate spectrograms or vocoder features frame by frame.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 501,
+ 506,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 501,
+ 506,
+ 515
+ ],
+ "score": 1.0,
+ "content": "Guo et al. (2019) suggest using an adversarial loss to reduce exposure bias (Bengio et al., 2015;",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 512,
+ 506,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 512,
+ 506,
+ 526
+ ],
+ "score": 1.0,
+ "content": "Ranzato et al., 2016) in such models. MelNet (Vasquez & Lewis, 2019) is autoregressive over",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 524,
+ 506,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 524,
+ 506,
+ 536
+ ],
+ "score": 1.0,
+ "content": "both time and frequency. ParaNet (Peng et al., 2019) and FastSpeech (Ren et al., 2019) are non-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 535,
+ 506,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 535,
+ 506,
+ 547
+ ],
+ "score": 1.0,
+ "content": "autoregressive, but they require distillation (Hinton et al., 2015) from an autoregressive model. Recent",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 546,
+ 506,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 546,
+ 506,
+ 558
+ ],
+ "score": 1.0,
+ "content": "flow-based approaches Flow-TTS (Miao et al., 2020) and Glow-TTS (Kim et al., 2020) are feed-",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 557,
+ 506,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 557,
+ 506,
+ 569
+ ],
+ "score": 1.0,
+ "content": "forward without requiring distillation. Most spectrogram generation models require training of a",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 568,
+ 505,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 568,
+ 505,
+ 580
+ ],
+ "score": 1.0,
+ "content": "custom vocoder model on generated spectrograms, because their predictions are imperfect and the",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 506,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 506,
+ 591
+ ],
+ "score": 1.0,
+ "content": "vocoder needs to be able to compensate for this2. Note that some of these works also propose new",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 590,
+ 386,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 590,
+ 386,
+ 602
+ ],
+ "score": 1.0,
+ "content": "vocoder architectures in tandem with spectrogram generation models.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 604,
+ 505,
+ 681
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "Unlike all of the aforementioned methods, as highlighted in Appendix J, Table 6, our model is a single",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "feed-forward neural network, trained end-to-end in a single stage, which produces waveforms given",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "character or phoneme sequences, and learns to align without additional supervision from auxiliary",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "sources (e.g. temporally aligned linguistic features from an external model) or teacher forcing. This",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 661
+ ],
+ "score": 1.0,
+ "content": "simplifies the training process considerably. Char2wav (Sotelo et al., 2017) is finetuned end-to-end",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 659,
+ 506,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 659,
+ 506,
+ 672
+ ],
+ "score": 1.0,
+ "content": "in the same fashion, but requires a pre-training stage with vocoder features used for intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 158,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 158,
+ 684
+ ],
+ "score": 1.0,
+ "content": "supervision.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 48
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 685,
+ 504,
+ 708
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 685,
+ 506,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 685,
+ 506,
+ 698
+ ],
+ "score": 1.0,
+ "content": "Spectrogram prediction losses have been used extensively for feed-forward audio prediction mod-",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 106,
+ 695,
+ 506,
+ 710
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 695,
+ 506,
+ 710
+ ],
+ "score": 1.0,
+ "content": "els (Yamamoto et al., 2019; 2020; Yang et al., 2020; Arık et al., 2018; Engel et al., 2020; Wang et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "index": 52.5
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 113,
+ 721,
+ 498,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 719,
+ 498,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 719,
+ 498,
+ 734
+ ],
+ "score": 1.0,
+ "content": "2This also implies that the spectrogram generation model and the vocoder have to be trained sequentially.",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 292,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "6",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 81,
+ 208,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 79,
+ 211,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 79,
+ 211,
+ 96
+ ],
+ "score": 1.0,
+ "content": "3 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 106,
+ 505,
+ 216
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 120
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 120
+ ],
+ "score": 1.0,
+ "content": "Speech generation saw significant quality improvements once treating it as a generative modelling",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 505,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 505,
+ 130
+ ],
+ "score": 1.0,
+ "content": "problem became the norm (Zen et al., 2009; van den Oord et al., 2016). Likelihood-based approaches",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 142
+ ],
+ "score": 1.0,
+ "content": "dominate, but generative adversarial networks (GANs) (Goodfellow et al., 2014) have been making",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 138,
+ 506,
+ 152
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 506,
+ 152
+ ],
+ "score": 1.0,
+ "content": "significant inroads recently. A common thread through most of the literature is a separation of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 150,
+ 506,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 150,
+ 506,
+ 163
+ ],
+ "score": 1.0,
+ "content": "the speech generation process into multiple stages: coarse-grained temporally aligned intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 161,
+ 506,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 161,
+ 506,
+ 174
+ ],
+ "score": 1.0,
+ "content": "representations, such as mel-spectrograms, are used to divide the task into more manageable sub-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 171,
+ 506,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 506,
+ 185
+ ],
+ "score": 1.0,
+ "content": "problems. Many works focus exclusively on either spectrogram generation or vocoding (generating a",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 183,
+ 506,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 183,
+ 506,
+ 195
+ ],
+ "score": 1.0,
+ "content": "waveform from a spectrogram). Our work is different in this respect, and we will point out which",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 194,
+ 505,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 194,
+ 505,
+ 207
+ ],
+ "score": 1.0,
+ "content": "stages of the generation process are addressed by each model. In Appendix J, Table 6 we compare",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 204,
+ 428,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 428,
+ 217
+ ],
+ "score": 1.0,
+ "content": "these methods in terms of the inputs and outputs to each stage of their pipelines.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 105,
+ 105,
+ 506,
+ 217
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 219,
+ 506,
+ 340
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 233
+ ],
+ "score": 1.0,
+ "content": "Initially, most likelihood-based models for TTS were autoregressive (van den Oord et al., 2016; Mehri",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 231,
+ 506,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 231,
+ 506,
+ 243
+ ],
+ "score": 1.0,
+ "content": "et al., 2017; Arik et al., 2017), which means that there is a sequential dependency between subsequent",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "time steps of the produced output signal. That makes these models impractical for real-time use,",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 507,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 507,
+ 266
+ ],
+ "score": 1.0,
+ "content": "although this can be addressed with careful engineering (Kalchbrenner et al., 2018; Valin & Skoglund,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 263,
+ 506,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 506,
+ 276
+ ],
+ "score": 1.0,
+ "content": "2019). More recently, flow-based models (Papamakarios et al., 2019) have been explored as a",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 274,
+ 506,
+ 287
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 506,
+ 287
+ ],
+ "score": 1.0,
+ "content": "feed-forward alternative that enables fast inference (without sequential dependencies). These can",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 285,
+ 506,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 285,
+ 506,
+ 298
+ ],
+ "score": 1.0,
+ "content": "either be trained directly using maximum likelihood (Prenger et al., 2019; Kim et al., 2019; Ping et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 295,
+ 507,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 295,
+ 507,
+ 310
+ ],
+ "score": 1.0,
+ "content": "2019b), or through distillation from an autoregressive model (van den Oord et al., 2018; Ping et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 307,
+ 506,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 307,
+ 506,
+ 320
+ ],
+ "score": 1.0,
+ "content": "2019a). All of these models produce waveforms conditioned on an intermediate representation: either",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 318,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "spectrograms or “linguistic features”, which contain temporally-aligned high-level information about",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 329,
+ 491,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 329,
+ 491,
+ 342
+ ],
+ "score": 1.0,
+ "content": "the speech signal. Spectrogram-conditioned waveform models are often referred to as vocoders.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 16,
+ "bbox_fs": [
+ 105,
+ 219,
+ 507,
+ 342
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 344,
+ 505,
+ 432
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 506,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 506,
+ 357
+ ],
+ "score": 1.0,
+ "content": "A growing body of work has applied GAN (Goodfellow et al., 2014) variants to speech synthe-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 355,
+ 506,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 355,
+ 506,
+ 367
+ ],
+ "score": 1.0,
+ "content": "sis (Donahue et al., 2019). An important advantage of adversarial losses for TTS is a focus on",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 366,
+ 506,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 506,
+ 378
+ ],
+ "score": 1.0,
+ "content": "realism over diversity; the latter is less important in this setting. This enables a more efficient use of",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 378,
+ 506,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 506,
+ 390
+ ],
+ "score": 1.0,
+ "content": "capacity compared to models trained with maximum likelihood. MelGAN (Kumar et al., 2019) and",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 387,
+ 505,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 505,
+ 401
+ ],
+ "score": 1.0,
+ "content": "Parallel WaveGAN (Yamamoto et al., 2020) are adversarial vocoders, producing raw waveforms from",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 398,
+ 507,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 398,
+ 507,
+ 413
+ ],
+ "score": 1.0,
+ "content": "mel-spectrograms. Neekhara et al. (2019) predict magnitude spectrograms from mel-spectrograms.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 410,
+ 505,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 505,
+ 422
+ ],
+ "score": 1.0,
+ "content": "Most directly related to our work is GAN-TTS (Binkowski et al., 2020), which produces waveforms ´",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 421,
+ 395,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 421,
+ 395,
+ 433
+ ],
+ "score": 1.0,
+ "content": "conditioned on aligned linguistic features, and we build upon that work.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 25.5,
+ "bbox_fs": [
+ 105,
+ 344,
+ 507,
+ 433
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 436,
+ 505,
+ 601
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 436,
+ 506,
+ 449
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 436,
+ 506,
+ 449
+ ],
+ "score": 1.0,
+ "content": "Another important line of work covers spectrogram generation from text. Such models rely on a",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 447,
+ 506,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 447,
+ 506,
+ 460
+ ],
+ "score": 1.0,
+ "content": "vocoder to convert the spectrograms into waveforms (for which one of the previously mentioned",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 458,
+ 506,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 458,
+ 506,
+ 471
+ ],
+ "score": 1.0,
+ "content": "models could be used, or a traditional spectrogram inversion technique (Griffin & Lim, 1984)).",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 467,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 467,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "Tacotron 1 & 2 (Wang et al., 2017; Shen et al., 2018), Deep Voice 2 & 3 (Gibiansky et al., 2017; Ping",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 104,
+ 479,
+ 506,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 479,
+ 506,
+ 494
+ ],
+ "score": 1.0,
+ "content": "et al., 2018), TransformerTTS (Li et al., 2019), Flowtron (Valle et al., 2020), and VoiceLoop (Taigman",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 491,
+ 507,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 491,
+ 507,
+ 504
+ ],
+ "score": 1.0,
+ "content": "et al., 2017) are autoregressive models that generate spectrograms or vocoder features frame by frame.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 501,
+ 506,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 501,
+ 506,
+ 515
+ ],
+ "score": 1.0,
+ "content": "Guo et al. (2019) suggest using an adversarial loss to reduce exposure bias (Bengio et al., 2015;",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 512,
+ 506,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 512,
+ 506,
+ 526
+ ],
+ "score": 1.0,
+ "content": "Ranzato et al., 2016) in such models. MelNet (Vasquez & Lewis, 2019) is autoregressive over",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 524,
+ 506,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 524,
+ 506,
+ 536
+ ],
+ "score": 1.0,
+ "content": "both time and frequency. ParaNet (Peng et al., 2019) and FastSpeech (Ren et al., 2019) are non-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 535,
+ 506,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 535,
+ 506,
+ 547
+ ],
+ "score": 1.0,
+ "content": "autoregressive, but they require distillation (Hinton et al., 2015) from an autoregressive model. Recent",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 546,
+ 506,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 546,
+ 506,
+ 558
+ ],
+ "score": 1.0,
+ "content": "flow-based approaches Flow-TTS (Miao et al., 2020) and Glow-TTS (Kim et al., 2020) are feed-",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 557,
+ 506,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 557,
+ 506,
+ 569
+ ],
+ "score": 1.0,
+ "content": "forward without requiring distillation. Most spectrogram generation models require training of a",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 568,
+ 505,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 568,
+ 505,
+ 580
+ ],
+ "score": 1.0,
+ "content": "custom vocoder model on generated spectrograms, because their predictions are imperfect and the",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 506,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 506,
+ 591
+ ],
+ "score": 1.0,
+ "content": "vocoder needs to be able to compensate for this2. Note that some of these works also propose new",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 590,
+ 386,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 590,
+ 386,
+ 602
+ ],
+ "score": 1.0,
+ "content": "vocoder architectures in tandem with spectrogram generation models.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 104,
+ 436,
+ 507,
+ 602
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 604,
+ 505,
+ 681
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "Unlike all of the aforementioned methods, as highlighted in Appendix J, Table 6, our model is a single",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "feed-forward neural network, trained end-to-end in a single stage, which produces waveforms given",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "character or phoneme sequences, and learns to align without additional supervision from auxiliary",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 638,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "sources (e.g. temporally aligned linguistic features from an external model) or teacher forcing. This",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 505,
+ 661
+ ],
+ "score": 1.0,
+ "content": "simplifies the training process considerably. Char2wav (Sotelo et al., 2017) is finetuned end-to-end",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 659,
+ 506,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 659,
+ 506,
+ 672
+ ],
+ "score": 1.0,
+ "content": "in the same fashion, but requires a pre-training stage with vocoder features used for intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 158,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 158,
+ 684
+ ],
+ "score": 1.0,
+ "content": "supervision.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 48,
+ "bbox_fs": [
+ 105,
+ 604,
+ 506,
+ 684
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 685,
+ 504,
+ 708
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 685,
+ 506,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 685,
+ 506,
+ 698
+ ],
+ "score": 1.0,
+ "content": "Spectrogram prediction losses have been used extensively for feed-forward audio prediction mod-",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 106,
+ 695,
+ 506,
+ 710
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 695,
+ 506,
+ 710
+ ],
+ "score": 1.0,
+ "content": "els (Yamamoto et al., 2019; 2020; Yang et al., 2020; Arık et al., 2018; Engel et al., 2020; Wang et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "index": 52.5,
+ "bbox_fs": [
+ 106,
+ 685,
+ 506,
+ 710
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 83,
+ 505,
+ 280
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 289,
+ 95
+ ],
+ "score": 1.0,
+ "content": "2019; Défossez et al., 2018). We note that the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 83,
+ 302,
+ 93
+ ],
+ "score": 0.88,
+ "content": "L _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 302,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "loss we use (along with (Défossez et al., 2018)), is",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": "comparatively simple, as spectrogram losses in the literature tend to have separate terms penalising",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "magnitudes, log-magnitudes and phase components, each with their own scaling factors, and often",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 129
+ ],
+ "score": 1.0,
+ "content": "across multiple resolutions. Dynamic time warping on spectrograms is a component of many speech",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 127,
+ 506,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 127,
+ 506,
+ 138
+ ],
+ "score": 1.0,
+ "content": "recognition systems (Sakoe, 1971; Sakoe & Chiba, 1978), and has also been used for evaluation of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 149
+ ],
+ "score": 1.0,
+ "content": "TTS systems (Sailor & Patil, 2014; Chevelu et al., 2015). Cuturi & Blondel (2017) proposed the soft",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 148,
+ 506,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 506,
+ 161
+ ],
+ "score": 1.0,
+ "content": "version of DTW we use in this work as a differentiable loss function for time series models. Kim",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 159,
+ 506,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 159,
+ 506,
+ 172
+ ],
+ "score": 1.0,
+ "content": "et al. (2020) propose Monotonic Alignment Search (MAS), which relates to DTW in that both use",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 171,
+ 506,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 506,
+ 183
+ ],
+ "score": 1.0,
+ "content": "dynamic programming to implicitly align sequences for TTS. However, they have different goals:",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 180,
+ 506,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 180,
+ 506,
+ 194
+ ],
+ "score": 1.0,
+ "content": "MAS finds the optimal alignment between the text and a latent representation, whereas we use",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 191,
+ 506,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 191,
+ 506,
+ 205
+ ],
+ "score": 1.0,
+ "content": "DTW to relax the constraints imposed by our spectrogram prediction loss term. Several mechanisms",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 203,
+ 506,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 203,
+ 506,
+ 216
+ ],
+ "score": 1.0,
+ "content": "have been proposed to exploit monotonicity in tasks that require sequence alignment, including",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 214,
+ 506,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 214,
+ 506,
+ 226
+ ],
+ "score": 1.0,
+ "content": "attention mechanisms (Graves, 2013; Zhang et al., 2018; Vasquez & Lewis, 2019; He et al., 2019;",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 237
+ ],
+ "score": 1.0,
+ "content": "Raffel et al., 2017; Chiu & Raffel, 2018), loss functions (Graves et al., 2006; Graves, 2012) and",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 248
+ ],
+ "score": 1.0,
+ "content": "search-based approaches (Kim et al., 2020). For TTS, incorporating this constraint has been found",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 246,
+ 506,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 506,
+ 261
+ ],
+ "score": 1.0,
+ "content": "to help generalisation to long sequences (Battenberg et al., 2020). We incorporate monotonicity by",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 257,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 257,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "using an interpolation mechanism, which is cheap to compute because it is not recurrent (unlike many",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 270,
+ 246,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 270,
+ 246,
+ 281
+ ],
+ "score": 1.0,
+ "content": "monotonic attention mechanisms).",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 8.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 294,
+ 192,
+ 306
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 292,
+ 195,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 292,
+ 195,
+ 309
+ ],
+ "score": 1.0,
+ "content": "4 EVALUATION",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 316,
+ 505,
+ 393
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 330
+ ],
+ "score": 1.0,
+ "content": "In this section we discuss the setup and results of our empirical evaluation, describing the hyperparam-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 327,
+ 505,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 327,
+ 505,
+ 340
+ ],
+ "score": 1.0,
+ "content": "eter settings used for training and validating the architectural decisions and loss function components",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 339,
+ 504,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 339,
+ 504,
+ 351
+ ],
+ "score": 1.0,
+ "content": "detailed in Section 2. Our primary metric used to evaluate speech quality is the Mean Opinion Score",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 350,
+ 505,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 350,
+ 505,
+ 362
+ ],
+ "score": 1.0,
+ "content": "(MOS) given by human raters, computed by taking the mean of 1-5 naturalness ratings given across",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 359,
+ 505,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 359,
+ 505,
+ 373
+ ],
+ "score": 1.0,
+ "content": "1000 held-out conditioning sequences. In Appendix I we also report the Fréchet DeepSpeech Distance",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 385
+ ],
+ "score": 1.0,
+ "content": "(FDSD), proposed by Binkowski et al. (2020) as a speech synthesis quality metric. Appendix A ´",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 414,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 414,
+ 395
+ ],
+ "score": 1.0,
+ "content": "reports training and evaluation hyperparameters we used for all experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 405,
+ 243,
+ 415
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 405,
+ 245,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 405,
+ 245,
+ 416
+ ],
+ "score": 1.0,
+ "content": "4.1 MULTI-SPEAKER DATASET",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 423,
+ 505,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 423,
+ 505,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 423,
+ 505,
+ 436
+ ],
+ "score": 1.0,
+ "content": "We train all models on a private dataset that consists of high-quality recordings of human speech",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 434,
+ 506,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 506,
+ 447
+ ],
+ "score": 1.0,
+ "content": "performed by professional voice actors, and corresponding text. The voice pool consists of 69 female",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 444,
+ 506,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 506,
+ 458
+ ],
+ "score": 1.0,
+ "content": "and male voices of North American English speakers, while the audio clips contain full sentences of",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 455,
+ 505,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 455,
+ 304,
+ 469
+ ],
+ "score": 1.0,
+ "content": "lengths varying from less than 1 to 20 seconds at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 456,
+ 335,
+ 467
+ ],
+ "score": 0.67,
+ "content": "2 4 \\mathrm { k H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 335,
+ 455,
+ 505,
+ 469
+ ],
+ "score": 1.0,
+ "content": "frequency. Individual voices are unevenly",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 467,
+ 505,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 467,
+ 505,
+ 480
+ ],
+ "score": 1.0,
+ "content": "distributed, accounting for from 15 minutes to over 51 hours of recorded speech, totalling 260.49",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 477,
+ 506,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 477,
+ 506,
+ 492
+ ],
+ "score": 1.0,
+ "content": "hours. At training time, we sample 2 second windows from the individual clips, post-padding those",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "shorter than 2 seconds with silence. For evaluation, we focus on the single most prolific speaker in",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 500,
+ 506,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 506,
+ 513
+ ],
+ "score": 1.0,
+ "content": "our dataset, with all our main MOS results reported with the model conditioned on that speaker ID,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 511,
+ 500,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 500,
+ 524
+ ],
+ "score": 1.0,
+ "content": "but also report MOS results for each of the top four speakers using our main multi-speaker model.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 533,
+ 170,
+ 544
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 171,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 171,
+ 546
+ ],
+ "score": 1.0,
+ "content": "4.2 RESULTS",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 551,
+ 505,
+ 651
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 551,
+ 506,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 551,
+ 506,
+ 564
+ ],
+ "score": 1.0,
+ "content": "In Table 1 we present quantitative results for our EATS model described in Section 2, as well as",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 562,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 562,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": "several ablations of the different model and learning signal components. The architecture and training",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 586
+ ],
+ "score": 1.0,
+ "content": "setup of each ablation is identical to our base EATS model except in terms of the differences described",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "by the columns in Table 1. Each ablation is “subtractive”, representing the full EATS system minus",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 609
+ ],
+ "score": 1.0,
+ "content": "one particular feature. Our main result achieved by the base multi-speaker model is a mean opinion",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "score (MOS) of 4.083. Although it is difficult to compare directly with prior results from the literature",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 631
+ ],
+ "score": 1.0,
+ "content": "due to dataset differences, we nonetheless include MOS results from prior works (Binkowski et al., ´",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 628,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 358,
+ 641
+ ],
+ "score": 1.0,
+ "content": "2020; van den Oord et al., 2016; 2018), with MOS in the 4.2 to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 359,
+ 629,
+ 379,
+ 640
+ ],
+ "score": 0.83,
+ "content": "4 . 4 +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 379,
+ 628,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "range. Compared to these prior",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 640,
+ 472,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 472,
+ 652
+ ],
+ "score": 1.0,
+ "content": "models, which rely on aligned linguistic features, EATS uses substantially less supervision.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 41
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 654,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "The No RWDs, No MelSpecD, and No Discriminators ablations all achieved substantially worse",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 679
+ ],
+ "score": 1.0,
+ "content": "MOS results than our proposed model, demonstrating the importance of adversarial feedback. In",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "particular, the No RWDs ablation, with an MOS of 2.526, demonstrates the importance of the raw",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "audio feedback, and removing RWDs significantly degrades the high frequency components. No",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "MelSpecD causes intermittent artifacts and distortion, and removing all discriminators results in",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 344,
+ 724
+ ],
+ "score": 1.0,
+ "content": "audio that sounds robotic and distorted throughout. The No",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 710,
+ 374,
+ 721
+ ],
+ "score": 0.86,
+ "content": "\\mathcal { L } _ { \\mathrm { l e n g t h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 709,
+ 406,
+ 724
+ ],
+ "score": 1.0,
+ "content": "and No",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 407,
+ 710,
+ 431,
+ 721
+ ],
+ "score": 0.68,
+ "content": "\\mathcal { L } _ { \\mathrm { p r e d } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 431,
+ 709,
+ 506,
+ 724
+ ],
+ "score": 1.0,
+ "content": "ablations result in",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "a model that does not train at all. Comparing our model with No DTW (MOS 3.559), the temporal",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "page_idx": 6,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 292,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 25,
+ 293,
+ 39
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 25,
+ 293,
+ 39
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 303,
+ 751,
+ 308,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 13,
+ "width": 7
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 83,
+ 505,
+ 280
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 289,
+ 95
+ ],
+ "score": 1.0,
+ "content": "2019; Défossez et al., 2018). We note that the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 83,
+ 302,
+ 93
+ ],
+ "score": 0.88,
+ "content": "L _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 302,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "loss we use (along with (Défossez et al., 2018)), is",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": "comparatively simple, as spectrogram losses in the literature tend to have separate terms penalising",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "magnitudes, log-magnitudes and phase components, each with their own scaling factors, and often",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 129
+ ],
+ "score": 1.0,
+ "content": "across multiple resolutions. Dynamic time warping on spectrograms is a component of many speech",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 127,
+ 506,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 127,
+ 506,
+ 138
+ ],
+ "score": 1.0,
+ "content": "recognition systems (Sakoe, 1971; Sakoe & Chiba, 1978), and has also been used for evaluation of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 149
+ ],
+ "score": 1.0,
+ "content": "TTS systems (Sailor & Patil, 2014; Chevelu et al., 2015). Cuturi & Blondel (2017) proposed the soft",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 148,
+ 506,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 148,
+ 506,
+ 161
+ ],
+ "score": 1.0,
+ "content": "version of DTW we use in this work as a differentiable loss function for time series models. Kim",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 159,
+ 506,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 159,
+ 506,
+ 172
+ ],
+ "score": 1.0,
+ "content": "et al. (2020) propose Monotonic Alignment Search (MAS), which relates to DTW in that both use",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 171,
+ 506,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 171,
+ 506,
+ 183
+ ],
+ "score": 1.0,
+ "content": "dynamic programming to implicitly align sequences for TTS. However, they have different goals:",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 180,
+ 506,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 180,
+ 506,
+ 194
+ ],
+ "score": 1.0,
+ "content": "MAS finds the optimal alignment between the text and a latent representation, whereas we use",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 191,
+ 506,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 191,
+ 506,
+ 205
+ ],
+ "score": 1.0,
+ "content": "DTW to relax the constraints imposed by our spectrogram prediction loss term. Several mechanisms",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 203,
+ 506,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 203,
+ 506,
+ 216
+ ],
+ "score": 1.0,
+ "content": "have been proposed to exploit monotonicity in tasks that require sequence alignment, including",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 214,
+ 506,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 214,
+ 506,
+ 226
+ ],
+ "score": 1.0,
+ "content": "attention mechanisms (Graves, 2013; Zhang et al., 2018; Vasquez & Lewis, 2019; He et al., 2019;",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 237
+ ],
+ "score": 1.0,
+ "content": "Raffel et al., 2017; Chiu & Raffel, 2018), loss functions (Graves et al., 2006; Graves, 2012) and",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 248
+ ],
+ "score": 1.0,
+ "content": "search-based approaches (Kim et al., 2020). For TTS, incorporating this constraint has been found",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 246,
+ 506,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 506,
+ 261
+ ],
+ "score": 1.0,
+ "content": "to help generalisation to long sequences (Battenberg et al., 2020). We incorporate monotonicity by",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 257,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 257,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "using an interpolation mechanism, which is cheap to compute because it is not recurrent (unlike many",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 270,
+ 246,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 270,
+ 246,
+ 281
+ ],
+ "score": 1.0,
+ "content": "monotonic attention mechanisms).",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 8.5,
+ "bbox_fs": [
+ 105,
+ 82,
+ 506,
+ 281
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 294,
+ 192,
+ 306
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 292,
+ 195,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 292,
+ 195,
+ 309
+ ],
+ "score": 1.0,
+ "content": "4 EVALUATION",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 316,
+ 505,
+ 393
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 330
+ ],
+ "score": 1.0,
+ "content": "In this section we discuss the setup and results of our empirical evaluation, describing the hyperparam-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 327,
+ 505,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 327,
+ 505,
+ 340
+ ],
+ "score": 1.0,
+ "content": "eter settings used for training and validating the architectural decisions and loss function components",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 339,
+ 504,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 339,
+ 504,
+ 351
+ ],
+ "score": 1.0,
+ "content": "detailed in Section 2. Our primary metric used to evaluate speech quality is the Mean Opinion Score",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 350,
+ 505,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 350,
+ 505,
+ 362
+ ],
+ "score": 1.0,
+ "content": "(MOS) given by human raters, computed by taking the mean of 1-5 naturalness ratings given across",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 359,
+ 505,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 359,
+ 505,
+ 373
+ ],
+ "score": 1.0,
+ "content": "1000 held-out conditioning sequences. In Appendix I we also report the Fréchet DeepSpeech Distance",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 385
+ ],
+ "score": 1.0,
+ "content": "(FDSD), proposed by Binkowski et al. (2020) as a speech synthesis quality metric. Appendix A ´",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 414,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 414,
+ 395
+ ],
+ "score": 1.0,
+ "content": "reports training and evaluation hyperparameters we used for all experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 104,
+ 315,
+ 506,
+ 395
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 405,
+ 243,
+ 415
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 405,
+ 245,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 405,
+ 245,
+ 416
+ ],
+ "score": 1.0,
+ "content": "4.1 MULTI-SPEAKER DATASET",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 423,
+ 505,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 423,
+ 505,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 423,
+ 505,
+ 436
+ ],
+ "score": 1.0,
+ "content": "We train all models on a private dataset that consists of high-quality recordings of human speech",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 434,
+ 506,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 506,
+ 447
+ ],
+ "score": 1.0,
+ "content": "performed by professional voice actors, and corresponding text. The voice pool consists of 69 female",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 444,
+ 506,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 506,
+ 458
+ ],
+ "score": 1.0,
+ "content": "and male voices of North American English speakers, while the audio clips contain full sentences of",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 455,
+ 505,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 455,
+ 304,
+ 469
+ ],
+ "score": 1.0,
+ "content": "lengths varying from less than 1 to 20 seconds at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 456,
+ 335,
+ 467
+ ],
+ "score": 0.67,
+ "content": "2 4 \\mathrm { k H z }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 335,
+ 455,
+ 505,
+ 469
+ ],
+ "score": 1.0,
+ "content": "frequency. Individual voices are unevenly",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 467,
+ 505,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 467,
+ 505,
+ 480
+ ],
+ "score": 1.0,
+ "content": "distributed, accounting for from 15 minutes to over 51 hours of recorded speech, totalling 260.49",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 477,
+ 506,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 477,
+ 506,
+ 492
+ ],
+ "score": 1.0,
+ "content": "hours. At training time, we sample 2 second windows from the individual clips, post-padding those",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "shorter than 2 seconds with silence. For evaluation, we focus on the single most prolific speaker in",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 500,
+ 506,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 506,
+ 513
+ ],
+ "score": 1.0,
+ "content": "our dataset, with all our main MOS results reported with the model conditioned on that speaker ID,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 511,
+ 500,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 500,
+ 524
+ ],
+ "score": 1.0,
+ "content": "but also report MOS results for each of the top four speakers using our main multi-speaker model.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 31,
+ "bbox_fs": [
+ 105,
+ 423,
+ 506,
+ 524
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 533,
+ 170,
+ 544
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 171,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 171,
+ 546
+ ],
+ "score": 1.0,
+ "content": "4.2 RESULTS",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 551,
+ 505,
+ 651
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 551,
+ 506,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 551,
+ 506,
+ 564
+ ],
+ "score": 1.0,
+ "content": "In Table 1 we present quantitative results for our EATS model described in Section 2, as well as",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 562,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 562,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": "several ablations of the different model and learning signal components. The architecture and training",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 586
+ ],
+ "score": 1.0,
+ "content": "setup of each ablation is identical to our base EATS model except in terms of the differences described",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "by the columns in Table 1. Each ablation is “subtractive”, representing the full EATS system minus",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 609
+ ],
+ "score": 1.0,
+ "content": "one particular feature. Our main result achieved by the base multi-speaker model is a mean opinion",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 606,
+ 506,
+ 620
+ ],
+ "score": 1.0,
+ "content": "score (MOS) of 4.083. Although it is difficult to compare directly with prior results from the literature",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 631
+ ],
+ "score": 1.0,
+ "content": "due to dataset differences, we nonetheless include MOS results from prior works (Binkowski et al., ´",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 628,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 358,
+ 641
+ ],
+ "score": 1.0,
+ "content": "2020; van den Oord et al., 2016; 2018), with MOS in the 4.2 to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 359,
+ 629,
+ 379,
+ 640
+ ],
+ "score": 0.83,
+ "content": "4 . 4 +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 379,
+ 628,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "range. Compared to these prior",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 640,
+ 472,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 472,
+ 652
+ ],
+ "score": 1.0,
+ "content": "models, which rely on aligned linguistic features, EATS uses substantially less supervision.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 41,
+ "bbox_fs": [
+ 105,
+ 551,
+ 506,
+ 652
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 654,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "The No RWDs, No MelSpecD, and No Discriminators ablations all achieved substantially worse",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 679
+ ],
+ "score": 1.0,
+ "content": "MOS results than our proposed model, demonstrating the importance of adversarial feedback. In",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "particular, the No RWDs ablation, with an MOS of 2.526, demonstrates the importance of the raw",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "audio feedback, and removing RWDs significantly degrades the high frequency components. No",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "MelSpecD causes intermittent artifacts and distortion, and removing all discriminators results in",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 344,
+ 724
+ ],
+ "score": 1.0,
+ "content": "audio that sounds robotic and distorted throughout. The No",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 710,
+ 374,
+ 721
+ ],
+ "score": 0.86,
+ "content": "\\mathcal { L } _ { \\mathrm { l e n g t h } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 709,
+ 406,
+ 724
+ ],
+ "score": 1.0,
+ "content": "and No",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 407,
+ 710,
+ 431,
+ 721
+ ],
+ "score": 0.68,
+ "content": "\\mathcal { L } _ { \\mathrm { p r e d } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 431,
+ 709,
+ 506,
+ 724
+ ],
+ "score": 1.0,
+ "content": "ablations result in",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "a model that does not train at all. Comparing our model with No DTW (MOS 3.559), the temporal",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 106,
+ 526,
+ 505,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 526,
+ 505,
+ 539
+ ],
+ "score": 1.0,
+ "content": "flexibility provided by dynamic time warping significantly improves fidelity: removing it causes",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 538,
+ 505,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 538,
+ 505,
+ 550
+ ],
+ "score": 1.0,
+ "content": "warbling and unnatural phoneme lengths. No Phonemes is trained with raw character inputs and",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 549,
+ 506,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 506,
+ 560
+ ],
+ "score": 1.0,
+ "content": "attains MOS 3.423, due to occasional mispronunciations and unusual stress patterns. No Mon. Int.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 559,
+ 506,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 559,
+ 506,
+ 573
+ ],
+ "score": 1.0,
+ "content": "uses an aligner with a transformer-based attention mechanism (described in Appendix G) in place",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 570,
+ 505,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 570,
+ 505,
+ 583
+ ],
+ "score": 1.0,
+ "content": "of our monotonic interpolation architecture, which turns out to generalise poorly to long utterances",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 581,
+ 505,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 581,
+ 505,
+ 594
+ ],
+ "score": 1.0,
+ "content": "(yielding MOS 3.551). Finally, comparing against training with only a Single Speaker (MOS 3.829)",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 506,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 506,
+ 605
+ ],
+ "score": 1.0,
+ "content": "shows that our EATS model benefits from a much larger multi-speaker dataset, even though MOS",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 603,
+ 506,
+ 616
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 506,
+ 616
+ ],
+ "score": 1.0,
+ "content": "is evaluated only on this same single speaker on which the ablation was solely trained. Samples",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 614,
+ 506,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 506,
+ 627
+ ],
+ "score": 1.0,
+ "content": "from each ablation are available at https://deepmind.com/research/publications/",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 626,
+ 335,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 626,
+ 335,
+ 636
+ ],
+ "score": 1.0,
+ "content": "End-to-End-Adversarial-Text-to-Speech.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 49,
+ "bbox_fs": [
+ 105,
+ 654,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 113,
+ 80,
+ 496,
+ 250
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 113,
+ 80,
+ 496,
+ 250
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 80,
+ 496,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 80,
+ 496,
+ 250
+ ],
+ "score": 0.985,
+ "html": " | Stages | 1 Stage | Notes |
| WaveNet (van den Oord et al.,2016) | AAu Ling | × | |
| SampleRNN (Mehri et al., 2017) | AAu | × | not a TTS model |
| Deep Voice (Arik et al.,2017) | Ch APhLing ARAu | × | uses segmentation model |
| WaveRNN (Kalchbrenner etal.,2018) | Ling AR ARAu | × | |
| LPCNet (Valin & Skoglund,2019) | ARAu Cep | × | |
| WaveGlow (Prenger et al.,2019) | MelS FF Au | × | |
| FloWaveNet (Kim et al.,2019) | FAu MelS | × | |
| WaveFlow (Ping et al.,2019b) | ARAu MelS | × | partially autoregressive |
| Par: WaveNet (van den Oord et al.,2018) | Ling FF* →Au | × | distillation |
| ClariNet (Ping et al.,2019a), teacher | ARAu Ch/Ph | √ | |
| ClariNet (Ping etal.,2019a),student | Au Ch/Ph | × | distillation |
| WaveGAN(Donahue et al.,2019) | Au | × | not a TTS model |
| MelGAN (Kumar etal., 2019) | MelsAu | × | |
| Par: WaveGAN(Yamamoto et al.,2020) | Ph AMels FAu | × | |
| AdVoc (Neekhara et al.,2019) | MelsMagS | × | |
| GAN-TTS (Binkowski etal., 2020) | Ling FAu FF | × | |
| Tacotron (Wang et al.,2017) | ChAMels MagS→Au | × | uses Griffin & Lim (1984) |
| Tacotron 2 (Shen et al.,2018) | AMelSAAu Ch | × | |
| Deep Voice 2(Gibiansky et al.,2017) | | × | uses segmentation model |
| DV2 Tacotron (Gibiansky et al.,2017) | ChAMagS AAu | × | |
| Deep Voice 3 (Ping et al.,2018) | Ch AMelS AAu | × | several alternative vocoders |
| TransformerTTS(Li etal.,2019) | Ch→Ph AMelS AAu | × | |
| Flowtron (Valle et al.,2020) | Ch AMelsAu | × | |
| VoiceLoop (Taigman etal.,2017) | Ph ALing → Au | × | |
| GAN Exposure (Guo et al.,2019) | Ph A MelS A Au | × | |
| MelNet (Vasquez & Lewis,2019) | AMelS→Au Ch- | × | |
| ParaNet (Peng et al.,2019) | FF* MelsAu Ch/Ph | × | distillation |
| FastSpeech (Ren et al.,2019) | FAu Ph FF* MelS | × | distillation |
| Flow-TTS (Miao et al.,2020) | Mels Au Ch | | |
| Glow-TTS(Kim et al.,2020) | PhMels Au | × × | |
| Char2wav (Sotelo et al.,2017) | ChALing ARAu | × | end-to-end finetuning |
| EATS (Ours) | Ch/Ph → Au | √ | |
",
+ "type": "table",
+ "image_path": "070d91881cd339f889f753d58ce44172edfccd9a0357addd1c45321e9cb41e33.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 110,
+ 137,
+ 496,
+ 276.66666666666663
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 110,
+ 276.66666666666663,
+ 496,
+ 416.33333333333326
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 110,
+ 416.33333333333326,
+ 496,
+ 555.9999999999999
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 561,
+ 506,
+ 672
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 562,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 562,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "Table 6: A comparison of TTS methods. The model stages described in each paper are shown",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 586
+ ],
+ "score": 1.0,
+ "content": "by linking together the inputs, outputs and intermediate representations that are used: characters",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 584,
+ 506,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 584,
+ 127,
+ 595
+ ],
+ "score": 0.43,
+ "content": "\\mathbf { \\Pi } ( \\mathbf { C h } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 127,
+ 584,
+ 177,
+ 597
+ ],
+ "score": 1.0,
+ "content": ", phonemes",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 177,
+ 584,
+ 195,
+ 595
+ ],
+ "score": 0.49,
+ "content": "\\mathbf { ( P h ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 196,
+ 584,
+ 506,
+ 597
+ ],
+ "score": 1.0,
+ "content": ", mel-spectrograms (MelS), magnitude spectrograms (MagS), cepstral fea-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 608
+ ],
+ "score": 1.0,
+ "content": "tures (Cep), linguistic features (Ling, such as phoneme durations and fundamental frequencies, or",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "WORLD (Morise et al., 2016) features for Char2wav (Sotelo et al., 2017) and VoiceLoop (Taigman",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 204,
+ 629
+ ],
+ "score": 1.0,
+ "content": "et al., 2017)), and audio",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 617,
+ 222,
+ 628
+ ],
+ "score": 0.39,
+ "content": "\\mathbf { \\Pi } ( \\mathbf { A u } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 223,
+ 617,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": ". Arrows with various superscripts describe model components: autore-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 104,
+ 627,
+ 507,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 627,
+ 401,
+ 641
+ ],
+ "score": 1.0,
+ "content": "gressive (AR), feed-forward (FF), or feed-forward requiring distillation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 402,
+ 628,
+ 426,
+ 639
+ ],
+ "score": 0.88,
+ "content": "( \\mathbf { F } \\mathbf { F } ^ { * } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 627,
+ 507,
+ 641
+ ],
+ "score": 1.0,
+ "content": ". Arrows without a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 639,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 639,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "superscript indicate components that do not require learning. 1 Stage means the model is trained in",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 650,
+ 506,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 506,
+ 663
+ ],
+ "score": 1.0,
+ "content": "a single stage to map from unaligned text/phonemes to audio (without, e.g., distillation or separate",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 449,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 449,
+ 673
+ ],
+ "score": 1.0,
+ "content": "vocoder training). EATS is the only feed-forward model that fulfills this requirement.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 7.5
+ }
+ ],
+ "index": 4.25
+ }
+ ],
+ "page_idx": 21,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 292,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "22",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 110,
+ 137,
+ 496,
+ 556
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 110,
+ 137,
+ 496,
+ 556
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 137,
+ 496,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 137,
+ 496,
+ 556
+ ],
+ "score": 0.982,
+ "html": " | Stages | 1 Stage | Notes |
| WaveNet (van den Oord et al.,2016) | AAu Ling | × | |
| SampleRNN (Mehri et al., 2017) | AAu | × | not a TTS model |
| Deep Voice (Arik et al.,2017) | Ch APhLing ARAu | × | uses segmentation model |
| WaveRNN (Kalchbrenner etal.,2018) | Ling AR ARAu | × | |
| LPCNet (Valin & Skoglund,2019) | ARAu Cep | × | |
| WaveGlow (Prenger et al.,2019) | MelS FF Au | × | |
| FloWaveNet (Kim et al.,2019) | FAu MelS | × | |
| WaveFlow (Ping et al.,2019b) | ARAu MelS | × | partially autoregressive |
| Par: WaveNet (van den Oord et al.,2018) | Ling FF* →Au | × | distillation |
| ClariNet (Ping et al.,2019a), teacher | ARAu Ch/Ph | √ | |
| ClariNet (Ping etal.,2019a),student | Au Ch/Ph | × | distillation |
| WaveGAN(Donahue et al.,2019) | Au | × | not a TTS model |
| MelGAN (Kumar etal., 2019) | MelsAu | × | |
| Par: WaveGAN(Yamamoto et al.,2020) | Ph AMels FAu | × | |
| AdVoc (Neekhara et al.,2019) | MelsMagS | × | |
| GAN-TTS (Binkowski etal., 2020) | Ling FAu FF | × | |
| Tacotron (Wang et al.,2017) | ChAMels MagS→Au | × | uses Griffin & Lim (1984) |
| Tacotron 2 (Shen et al.,2018) | AMelSAAu Ch | × | |
| Deep Voice 2(Gibiansky et al.,2017) | | × | uses segmentation model |
| DV2 Tacotron (Gibiansky et al.,2017) | ChAMagS AAu | × | |
| Deep Voice 3 (Ping et al.,2018) | Ch AMelS AAu | × | several alternative vocoders |
| TransformerTTS(Li etal.,2019) | Ch→Ph AMelS AAu | × | |
| Flowtron (Valle et al.,2020) | Ch AMelsAu | × | |
| VoiceLoop (Taigman etal.,2017) | Ph ALing → Au | × | |
| GAN Exposure (Guo et al.,2019) | Ph A MelS A Au | × | |
| MelNet (Vasquez & Lewis,2019) | AMelS→Au Ch- | × | |
| ParaNet (Peng et al.,2019) | FF* MelsAu Ch/Ph | × | distillation |
| FastSpeech (Ren et al.,2019) | FAu Ph FF* MelS | × | distillation |
| Flow-TTS (Miao et al.,2020) | Mels Au Ch | | |
| Glow-TTS(Kim et al.,2020) | PhMels Au | × × | |
| Char2wav (Sotelo et al.,2017) | ChALing ARAu | × | end-to-end finetuning |
| EATS (Ours) | Ch/Ph → Au | √ | |
",
+ "type": "table",
+ "image_path": "070d91881cd339f889f753d58ce44172edfccd9a0357addd1c45321e9cb41e33.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 110,
+ 137,
+ 496,
+ 276.66666666666663
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 110,
+ 276.66666666666663,
+ 496,
+ 416.33333333333326
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 110,
+ 416.33333333333326,
+ 496,
+ 555.9999999999999
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 561,
+ 506,
+ 672
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 562,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 562,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "Table 6: A comparison of TTS methods. The model stages described in each paper are shown",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 586
+ ],
+ "score": 1.0,
+ "content": "by linking together the inputs, outputs and intermediate representations that are used: characters",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 584,
+ 506,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 584,
+ 127,
+ 595
+ ],
+ "score": 0.43,
+ "content": "\\mathbf { \\Pi } ( \\mathbf { C h } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 127,
+ 584,
+ 177,
+ 597
+ ],
+ "score": 1.0,
+ "content": ", phonemes",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 177,
+ 584,
+ 195,
+ 595
+ ],
+ "score": 0.49,
+ "content": "\\mathbf { ( P h ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 196,
+ 584,
+ 506,
+ 597
+ ],
+ "score": 1.0,
+ "content": ", mel-spectrograms (MelS), magnitude spectrograms (MagS), cepstral fea-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 608
+ ],
+ "score": 1.0,
+ "content": "tures (Cep), linguistic features (Ling, such as phoneme durations and fundamental frequencies, or",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "WORLD (Morise et al., 2016) features for Char2wav (Sotelo et al., 2017) and VoiceLoop (Taigman",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 204,
+ 629
+ ],
+ "score": 1.0,
+ "content": "et al., 2017)), and audio",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 617,
+ 222,
+ 628
+ ],
+ "score": 0.39,
+ "content": "\\mathbf { \\Pi } ( \\mathbf { A u } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 223,
+ 617,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": ". Arrows with various superscripts describe model components: autore-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 104,
+ 627,
+ 507,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 627,
+ 401,
+ 641
+ ],
+ "score": 1.0,
+ "content": "gressive (AR), feed-forward (FF), or feed-forward requiring distillation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 402,
+ 628,
+ 426,
+ 639
+ ],
+ "score": 0.88,
+ "content": "( \\mathbf { F } \\mathbf { F } ^ { * } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 627,
+ 507,
+ 641
+ ],
+ "score": 1.0,
+ "content": ". Arrows without a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 639,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 639,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "superscript indicate components that do not require learning. 1 Stage means the model is trained in",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 650,
+ 506,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 506,
+ 663
+ ],
+ "score": 1.0,
+ "content": "a single stage to map from unaligned text/phonemes to audio (without, e.g., distillation or separate",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 449,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 449,
+ 673
+ ],
+ "score": 1.0,
+ "content": "vocoder training). EATS is the only feed-forward model that fulfills this requirement.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 7.5
+ }
+ ],
+ "index": 4.25
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 81,
+ 289,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 79,
+ 292,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 79,
+ 292,
+ 97
+ ],
+ "score": 1.0,
+ "content": "J COMPARISON OF TTS METHODS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 104,
+ 505,
+ 181
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "In Table 6 we compare recent TTS approaches in terms of the inputs and outputs to each stage of",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 114,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 114,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "the pipeline, and whether they are learnt in a single stage or multiple stages. Differentiating EATS",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": "from each prior approach is the fact that it learns a feed-forward mapping from text/phonemes to",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 150
+ ],
+ "score": 1.0,
+ "content": "audio end-to-end in a single stage, without requiring distillation or separate vocoder training. The",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 146,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 146,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "ClariNet teacher model (Ping et al., 2019a) is also trained in a single stage, but it uses teacher forcing",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "to achieve this, requiring the model to be autoregressive. A separate distillation stage is necessary to",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 169,
+ 272,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 272,
+ 183
+ ],
+ "score": 1.0,
+ "content": "obtain a feed-forward model in this case.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "page_idx": 22,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 292,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 298,
+ 750,
+ 313,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 298,
+ 750,
+ 313,
+ 764
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 15
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 81,
+ 289,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 79,
+ 292,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 79,
+ 292,
+ 97
+ ],
+ "score": 1.0,
+ "content": "J COMPARISON OF TTS METHODS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 104,
+ 505,
+ 181
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "In Table 6 we compare recent TTS approaches in terms of the inputs and outputs to each stage of",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 114,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 114,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "the pipeline, and whether they are learnt in a single stage or multiple stages. Differentiating EATS",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": "from each prior approach is the fact that it learns a feed-forward mapping from text/phonemes to",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 505,
+ 150
+ ],
+ "score": 1.0,
+ "content": "audio end-to-end in a single stage, without requiring distillation or separate vocoder training. The",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 146,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 146,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "ClariNet teacher model (Ping et al., 2019a) is also trained in a single stage, but it uses teacher forcing",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "to achieve this, requiring the model to be autoregressive. A separate distillation stage is necessary to",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 169,
+ 272,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 272,
+ 183
+ ],
+ "score": 1.0,
+ "content": "obtain a feed-forward model in this case.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 105,
+ 104,
+ 506,
+ 183
+ ]
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file
diff --git a/parse/train/rsf1z-JSj87/rsf1z-JSj87_model.json b/parse/train/rsf1z-JSj87/rsf1z-JSj87_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..d6fec7845e847986f23d99b9c6d561ef4162a040
--- /dev/null
+++ b/parse/train/rsf1z-JSj87/rsf1z-JSj87_model.json
@@ -0,0 +1,28971 @@
+[
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1149,
+ 1404,
+ 1149,
+ 1404,
+ 1423,
+ 297,
+ 1423
+ ],
+ "score": 0.981
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 397,
+ 558,
+ 1302,
+ 558,
+ 1302,
+ 1014,
+ 397,
+ 1014
+ ],
+ "score": 0.98
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1433,
+ 1403,
+ 1433,
+ 1403,
+ 1648,
+ 298,
+ 1648
+ ],
+ "score": 0.977
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1658,
+ 1403,
+ 1658,
+ 1403,
+ 1843,
+ 298,
+ 1843
+ ],
+ "score": 0.972
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 314,
+ 313,
+ 1331,
+ 313,
+ 1331,
+ 408,
+ 314,
+ 408
+ ],
+ "score": 0.97
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 297,
+ 1947,
+ 1404,
+ 1947,
+ 1404,
+ 2033,
+ 297,
+ 2033
+ ],
+ "score": 0.957
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 300,
+ 1852,
+ 1402,
+ 1852,
+ 1402,
+ 1915,
+ 300,
+ 1915
+ ],
+ "score": 0.935
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 296,
+ 219,
+ 1265,
+ 219,
+ 1265,
+ 269,
+ 296,
+ 269
+ ],
+ "score": 0.928
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 302,
+ 1080,
+ 573,
+ 1080,
+ 573,
+ 1115,
+ 302,
+ 1115
+ ],
+ "score": 0.893
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 299,
+ 75,
+ 813,
+ 75,
+ 813,
+ 104,
+ 299,
+ 104
+ ],
+ "score": 0.88
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 773,
+ 488,
+ 926,
+ 488,
+ 926,
+ 521,
+ 773,
+ 521
+ ],
+ "score": 0.845
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 856,
+ 2088,
+ 856,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.753
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1212,
+ 1690,
+ 1309,
+ 1690,
+ 1309,
+ 1720,
+ 1212,
+ 1720
+ ],
+ "score": 0.73,
+ "latex": "( 2 0 0 \\ : \\mathrm { H z } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 924,
+ 1812,
+ 1010,
+ 1812,
+ 1010,
+ 1841,
+ 924,
+ 1841
+ ],
+ "score": 0.7,
+ "latex": "2 4 \\mathrm { k H z }"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 329.0,
+ 1943.0,
+ 915.0,
+ 1943.0,
+ 915.0,
+ 1980.0,
+ 329.0,
+ 1980.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 358.0,
+ 1974.0,
+ 1404.0,
+ 1974.0,
+ 1404.0,
+ 2010.0,
+ 358.0,
+ 2010.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 2005.0,
+ 1060.0,
+ 2005.0,
+ 1060.0,
+ 2035.0,
+ 292.0,
+ 2035.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 219.0,
+ 1267.0,
+ 219.0,
+ 1267.0,
+ 274.0,
+ 295.0,
+ 274.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1077.0,
+ 579.0,
+ 1077.0,
+ 579.0,
+ 1124.0,
+ 294.0,
+ 1124.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 72.0,
+ 815.0,
+ 72.0,
+ 815.0,
+ 108.0,
+ 295.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 769.0,
+ 486.0,
+ 932.0,
+ 486.0,
+ 932.0,
+ 525.0,
+ 769.0,
+ 525.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 840.0,
+ 2088.0,
+ 857.0,
+ 2088.0,
+ 857.0,
+ 2116.0,
+ 840.0,
+ 2116.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1148.0,
+ 1404.0,
+ 1148.0,
+ 1404.0,
+ 1185.0,
+ 294.0,
+ 1185.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1180.0,
+ 1404.0,
+ 1180.0,
+ 1404.0,
+ 1215.0,
+ 293.0,
+ 1215.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 290.0,
+ 1209.0,
+ 1405.0,
+ 1209.0,
+ 1405.0,
+ 1247.0,
+ 290.0,
+ 1247.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1241.0,
+ 1405.0,
+ 1241.0,
+ 1405.0,
+ 1277.0,
+ 294.0,
+ 1277.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1269.0,
+ 1407.0,
+ 1269.0,
+ 1407.0,
+ 1309.0,
+ 292.0,
+ 1309.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1299.0,
+ 1404.0,
+ 1299.0,
+ 1404.0,
+ 1339.0,
+ 292.0,
+ 1339.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1330.0,
+ 1402.0,
+ 1330.0,
+ 1402.0,
+ 1367.0,
+ 294.0,
+ 1367.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1359.0,
+ 1406.0,
+ 1359.0,
+ 1406.0,
+ 1403.0,
+ 292.0,
+ 1403.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1389.0,
+ 1362.0,
+ 1389.0,
+ 1362.0,
+ 1431.0,
+ 292.0,
+ 1431.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 392.0,
+ 554.0,
+ 1306.0,
+ 554.0,
+ 1306.0,
+ 596.0,
+ 392.0,
+ 596.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 589.0,
+ 1306.0,
+ 589.0,
+ 1306.0,
+ 622.0,
+ 393.0,
+ 622.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 620.0,
+ 1305.0,
+ 620.0,
+ 1305.0,
+ 652.0,
+ 394.0,
+ 652.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 392.0,
+ 649.0,
+ 1306.0,
+ 649.0,
+ 1306.0,
+ 683.0,
+ 392.0,
+ 683.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 681.0,
+ 1306.0,
+ 681.0,
+ 1306.0,
+ 714.0,
+ 394.0,
+ 714.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 710.0,
+ 1306.0,
+ 710.0,
+ 1306.0,
+ 745.0,
+ 393.0,
+ 745.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 741.0,
+ 1306.0,
+ 741.0,
+ 1306.0,
+ 776.0,
+ 393.0,
+ 776.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 771.0,
+ 1305.0,
+ 771.0,
+ 1305.0,
+ 805.0,
+ 393.0,
+ 805.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 800.0,
+ 1306.0,
+ 800.0,
+ 1306.0,
+ 837.0,
+ 393.0,
+ 837.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 392.0,
+ 831.0,
+ 1308.0,
+ 831.0,
+ 1308.0,
+ 868.0,
+ 392.0,
+ 868.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 392.0,
+ 861.0,
+ 1305.0,
+ 861.0,
+ 1305.0,
+ 898.0,
+ 392.0,
+ 898.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 396.0,
+ 894.0,
+ 1305.0,
+ 894.0,
+ 1305.0,
+ 926.0,
+ 396.0,
+ 926.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 924.0,
+ 1306.0,
+ 924.0,
+ 1306.0,
+ 959.0,
+ 393.0,
+ 959.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 392.0,
+ 953.0,
+ 1306.0,
+ 953.0,
+ 1306.0,
+ 989.0,
+ 392.0,
+ 989.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 396.0,
+ 985.0,
+ 664.0,
+ 985.0,
+ 664.0,
+ 1018.0,
+ 396.0,
+ 1018.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1431.0,
+ 1404.0,
+ 1431.0,
+ 1404.0,
+ 1472.0,
+ 292.0,
+ 1472.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1464.0,
+ 1406.0,
+ 1464.0,
+ 1406.0,
+ 1499.0,
+ 292.0,
+ 1499.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1496.0,
+ 1405.0,
+ 1496.0,
+ 1405.0,
+ 1531.0,
+ 294.0,
+ 1531.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1527.0,
+ 1405.0,
+ 1527.0,
+ 1405.0,
+ 1560.0,
+ 293.0,
+ 1560.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1555.0,
+ 1407.0,
+ 1555.0,
+ 1407.0,
+ 1592.0,
+ 292.0,
+ 1592.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1587.0,
+ 1405.0,
+ 1587.0,
+ 1405.0,
+ 1622.0,
+ 294.0,
+ 1622.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1618.0,
+ 572.0,
+ 1618.0,
+ 572.0,
+ 1652.0,
+ 296.0,
+ 1652.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1658.0,
+ 1405.0,
+ 1658.0,
+ 1405.0,
+ 1693.0,
+ 296.0,
+ 1693.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1689.0,
+ 1211.0,
+ 1689.0,
+ 1211.0,
+ 1725.0,
+ 294.0,
+ 1725.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1310.0,
+ 1689.0,
+ 1405.0,
+ 1689.0,
+ 1405.0,
+ 1725.0,
+ 1310.0,
+ 1725.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1721.0,
+ 1405.0,
+ 1721.0,
+ 1405.0,
+ 1753.0,
+ 296.0,
+ 1753.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1748.0,
+ 1409.0,
+ 1748.0,
+ 1409.0,
+ 1791.0,
+ 292.0,
+ 1791.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1780.0,
+ 1407.0,
+ 1780.0,
+ 1407.0,
+ 1817.0,
+ 292.0,
+ 1817.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1810.0,
+ 923.0,
+ 1810.0,
+ 923.0,
+ 1846.0,
+ 294.0,
+ 1846.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1011.0,
+ 1810.0,
+ 1215.0,
+ 1810.0,
+ 1215.0,
+ 1846.0,
+ 1011.0,
+ 1846.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 309.0,
+ 312.0,
+ 1334.0,
+ 312.0,
+ 1334.0,
+ 352.0,
+ 309.0,
+ 352.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 311.0,
+ 345.0,
+ 441.0,
+ 345.0,
+ 441.0,
+ 380.0,
+ 311.0,
+ 380.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 313.0,
+ 374.0,
+ 1236.0,
+ 374.0,
+ 1236.0,
+ 414.0,
+ 313.0,
+ 414.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1852.0,
+ 1404.0,
+ 1852.0,
+ 1404.0,
+ 1887.0,
+ 296.0,
+ 1887.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 298.0,
+ 1884.0,
+ 1404.0,
+ 1884.0,
+ 1404.0,
+ 1915.0,
+ 298.0,
+ 1915.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 0,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 976,
+ 1404,
+ 976,
+ 1404,
+ 1254,
+ 297,
+ 1254
+ ],
+ "score": 0.984
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 630,
+ 1405,
+ 630,
+ 1405,
+ 968,
+ 297,
+ 968
+ ],
+ "score": 0.983
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1696,
+ 1406,
+ 1696,
+ 1406,
+ 2035,
+ 297,
+ 2035
+ ],
+ "score": 0.982
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1364,
+ 1405,
+ 1364,
+ 1405,
+ 1611,
+ 297,
+ 1611
+ ],
+ "score": 0.982
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 313,
+ 1406,
+ 313,
+ 1406,
+ 526,
+ 297,
+ 526
+ ],
+ "score": 0.954
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 293,
+ 229,
+ 1403,
+ 229,
+ 1403,
+ 292,
+ 293,
+ 292
+ ],
+ "score": 0.936
+ },
+ {
+ "category_id": 8,
+ "poly": [
+ 568,
+ 1311,
+ 1127,
+ 1311,
+ 1127,
+ 1353,
+ 568,
+ 1353
+ ],
+ "score": 0.928
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1263,
+ 967,
+ 1263,
+ 967,
+ 1295,
+ 297,
+ 1295
+ ],
+ "score": 0.919
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 299,
+ 565,
+ 481,
+ 565,
+ 481,
+ 601,
+ 299,
+ 601
+ ],
+ "score": 0.902
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 300,
+ 75,
+ 813,
+ 75,
+ 813,
+ 105,
+ 300,
+ 105
+ ],
+ "score": 0.901
+ },
+ {
+ "category_id": 9,
+ "poly": [
+ 1367,
+ 1316,
+ 1400,
+ 1316,
+ 1400,
+ 1345,
+ 1367,
+ 1345
+ ],
+ "score": 0.897
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 299,
+ 1645,
+ 478,
+ 1645,
+ 478,
+ 1676,
+ 299,
+ 1676
+ ],
+ "score": 0.887
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 858,
+ 2088,
+ 858,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.724
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 859,
+ 2088,
+ 859,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.109
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1148,
+ 1910,
+ 1317,
+ 1910,
+ 1317,
+ 1948,
+ 1148,
+ 1948
+ ],
+ "score": 0.93,
+ "latex": "\\begin{array} { r } { c _ { n } = e _ { n } - \\frac { 1 } { 2 } l _ { n } } \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 855,
+ 1972,
+ 970,
+ 1972,
+ 970,
+ 2006,
+ 855,
+ 2006
+ ],
+ "score": 0.93,
+ "latex": "S = \\lceil e _ { N } \\rceil"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 1728,
+ 461,
+ 1728,
+ 461,
+ 1762,
+ 298,
+ 1762
+ ],
+ "score": 0.92,
+ "latex": "{ \\bf h } = f ( { \\bf x } , { \\bf z } , { \\bf s } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 587,
+ 1910,
+ 766,
+ 1910,
+ 766,
+ 1946,
+ 587,
+ 1946
+ ],
+ "score": 0.92,
+ "latex": "\\begin{array} { r } { e _ { n } = \\sum _ { m = 1 } ^ { n } l _ { m } } \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 1850,
+ 480,
+ 1850,
+ 480,
+ 1884,
+ 298,
+ 1884
+ ],
+ "score": 0.92,
+ "latex": "l _ { n } = g ( h _ { n } , \\mathbf { z } , \\mathbf { s } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 580,
+ 1698,
+ 793,
+ 1698,
+ 793,
+ 1731,
+ 580,
+ 1731
+ ],
+ "score": 0.91,
+ "latex": "{ \\bf x } = ( x _ { 1 } , \\dots , x _ { N } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 374,
+ 1367,
+ 457,
+ 1367,
+ 457,
+ 1399,
+ 374,
+ 1399
+ ],
+ "score": 0.91,
+ "latex": "\\mathcal { L } _ { G , \\mathrm { a d v } }"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 569,
+ 1310,
+ 1127,
+ 1310,
+ 1127,
+ 1353,
+ 569,
+ 1353
+ ],
+ "score": 0.89,
+ "latex": "\\mathcal { L } _ { G } = \\mathcal { L } _ { G , \\mathrm { a d v } } + \\lambda _ { \\mathrm { p r e d } } \\cdot \\mathcal { L } _ { \\mathrm { p r e d } } ^ { \\prime \\prime } + \\lambda _ { \\mathrm { l e n g t h } } \\cdot \\mathcal { L } _ { \\mathrm { l e n g t h } } ,"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1134,
+ 2004,
+ 1168,
+ 2004,
+ 1168,
+ 2034,
+ 1134,
+ 2034
+ ],
+ "score": 0.88,
+ "latex": "h _ { n }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 982,
+ 1549,
+ 1065,
+ 1549,
+ 1065,
+ 1582,
+ 982,
+ 1582
+ ],
+ "score": 0.88,
+ "latex": "( \\mathcal { L } _ { \\mathrm { { p r e d } } } ^ { \\prime \\prime } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1104,
+ 907,
+ 1171,
+ 907,
+ 1171,
+ 935,
+ 1104,
+ 935
+ ],
+ "score": 0.88,
+ "latex": "2 0 0 \\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 549,
+ 1730,
+ 569,
+ 1730,
+ 569,
+ 1761,
+ 549,
+ 1761
+ ],
+ "score": 0.86,
+ "latex": "f"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 576,
+ 1973,
+ 770,
+ 1973,
+ 770,
+ 2006,
+ 576,
+ 2006
+ ],
+ "score": 0.86,
+ "latex": "\\mathbf { a } = ( a _ { 1 } , \\ldots , a _ { S } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 431,
+ 2008,
+ 458,
+ 2008,
+ 458,
+ 2034,
+ 431,
+ 2034
+ ],
+ "score": 0.84,
+ "latex": "a _ { t }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 600,
+ 1194,
+ 620,
+ 1194,
+ 620,
+ 1223,
+ 600,
+ 1223
+ ],
+ "score": 0.83,
+ "latex": "\\mu"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 912,
+ 1700,
+ 940,
+ 1700,
+ 940,
+ 1726,
+ 912,
+ 1726
+ ],
+ "score": 0.8,
+ "latex": "N"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 569,
+ 1855,
+ 588,
+ 1855,
+ 588,
+ 1883,
+ 569,
+ 1883
+ ],
+ "score": 0.79,
+ "latex": "g"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 1227,
+ 317,
+ 1227,
+ 317,
+ 1254,
+ 298,
+ 1254
+ ],
+ "score": 0.75,
+ "latex": "\\mu"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1199,
+ 1550,
+ 1296,
+ 1550,
+ 1296,
+ 1583,
+ 1199,
+ 1583
+ ],
+ "score": 0.75,
+ "latex": "( \\mathcal { L } _ { \\mathrm { l e n g t h } } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1312,
+ 784,
+ 1396,
+ 784,
+ 1396,
+ 813,
+ 1312,
+ 813
+ ],
+ "score": 0.69,
+ "latex": "2 0 0 \\mathrm { H z }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 629,
+ 662,
+ 716,
+ 662,
+ 716,
+ 692,
+ 629,
+ 692
+ ],
+ "score": 0.67,
+ "latex": "2 4 \\mathrm { k H z }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 618,
+ 1106,
+ 636,
+ 1106,
+ 636,
+ 1128,
+ 618,
+ 1128
+ ],
+ "score": 0.56,
+ "latex": "\\mathbf { z }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1244,
+ 1765,
+ 1263,
+ 1765,
+ 1263,
+ 1787,
+ 1244,
+ 1787
+ ],
+ "score": 0.52,
+ "latex": "\\mathbf { z }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 483,
+ 1973,
+ 565,
+ 1973,
+ 565,
+ 2002,
+ 483,
+ 2002
+ ],
+ "score": 0.52,
+ "latex": "2 0 0 \\mathrm { H z }"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 559.0,
+ 486.0,
+ 559.0,
+ 486.0,
+ 609.0,
+ 292.0,
+ 609.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 73.0,
+ 815.0,
+ 73.0,
+ 815.0,
+ 108.0,
+ 295.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1642.0,
+ 482.0,
+ 1642.0,
+ 482.0,
+ 1681.0,
+ 294.0,
+ 1681.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 839.0,
+ 2085.0,
+ 862.0,
+ 2085.0,
+ 862.0,
+ 2120.0,
+ 839.0,
+ 2120.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 838.0,
+ 2085.0,
+ 862.0,
+ 2085.0,
+ 862.0,
+ 2120.0,
+ 838.0,
+ 2120.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 976.0,
+ 1407.0,
+ 976.0,
+ 1407.0,
+ 1013.0,
+ 295.0,
+ 1013.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1007.0,
+ 1406.0,
+ 1007.0,
+ 1406.0,
+ 1044.0,
+ 295.0,
+ 1044.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1035.0,
+ 1406.0,
+ 1035.0,
+ 1406.0,
+ 1076.0,
+ 292.0,
+ 1076.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1066.0,
+ 1404.0,
+ 1066.0,
+ 1404.0,
+ 1107.0,
+ 294.0,
+ 1107.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1098.0,
+ 617.0,
+ 1098.0,
+ 617.0,
+ 1135.0,
+ 292.0,
+ 1135.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 637.0,
+ 1098.0,
+ 1404.0,
+ 1098.0,
+ 1404.0,
+ 1135.0,
+ 637.0,
+ 1135.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1131.0,
+ 1404.0,
+ 1131.0,
+ 1404.0,
+ 1165.0,
+ 293.0,
+ 1165.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1158.0,
+ 1405.0,
+ 1158.0,
+ 1405.0,
+ 1196.0,
+ 292.0,
+ 1196.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1190.0,
+ 599.0,
+ 1190.0,
+ 599.0,
+ 1227.0,
+ 295.0,
+ 1227.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 621.0,
+ 1190.0,
+ 1406.0,
+ 1190.0,
+ 1406.0,
+ 1227.0,
+ 621.0,
+ 1227.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1219.0,
+ 297.0,
+ 1219.0,
+ 297.0,
+ 1259.0,
+ 294.0,
+ 1259.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 318.0,
+ 1219.0,
+ 1240.0,
+ 1219.0,
+ 1240.0,
+ 1259.0,
+ 318.0,
+ 1259.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 630.0,
+ 1405.0,
+ 630.0,
+ 1405.0,
+ 667.0,
+ 295.0,
+ 667.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 661.0,
+ 628.0,
+ 661.0,
+ 628.0,
+ 699.0,
+ 294.0,
+ 699.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 717.0,
+ 661.0,
+ 1406.0,
+ 661.0,
+ 1406.0,
+ 699.0,
+ 717.0,
+ 699.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 693.0,
+ 1406.0,
+ 693.0,
+ 1406.0,
+ 728.0,
+ 295.0,
+ 728.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 723.0,
+ 1405.0,
+ 723.0,
+ 1405.0,
+ 758.0,
+ 294.0,
+ 758.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 752.0,
+ 1405.0,
+ 752.0,
+ 1405.0,
+ 789.0,
+ 294.0,
+ 789.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 784.0,
+ 1311.0,
+ 784.0,
+ 1311.0,
+ 819.0,
+ 295.0,
+ 819.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1397.0,
+ 784.0,
+ 1407.0,
+ 784.0,
+ 1407.0,
+ 819.0,
+ 1397.0,
+ 819.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 814.0,
+ 1405.0,
+ 814.0,
+ 1405.0,
+ 850.0,
+ 294.0,
+ 850.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 844.0,
+ 1407.0,
+ 844.0,
+ 1407.0,
+ 880.0,
+ 292.0,
+ 880.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 877.0,
+ 1405.0,
+ 877.0,
+ 1405.0,
+ 908.0,
+ 296.0,
+ 908.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 905.0,
+ 1103.0,
+ 905.0,
+ 1103.0,
+ 941.0,
+ 292.0,
+ 941.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1172.0,
+ 905.0,
+ 1405.0,
+ 905.0,
+ 1405.0,
+ 941.0,
+ 1172.0,
+ 941.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 938.0,
+ 1301.0,
+ 938.0,
+ 1301.0,
+ 969.0,
+ 295.0,
+ 969.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1695.0,
+ 579.0,
+ 1695.0,
+ 579.0,
+ 1736.0,
+ 294.0,
+ 1736.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 794.0,
+ 1695.0,
+ 911.0,
+ 1695.0,
+ 911.0,
+ 1736.0,
+ 794.0,
+ 1736.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 941.0,
+ 1695.0,
+ 1406.0,
+ 1695.0,
+ 1406.0,
+ 1736.0,
+ 941.0,
+ 1736.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1728.0,
+ 297.0,
+ 1728.0,
+ 297.0,
+ 1763.0,
+ 294.0,
+ 1763.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 462.0,
+ 1728.0,
+ 548.0,
+ 1728.0,
+ 548.0,
+ 1763.0,
+ 462.0,
+ 1763.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 570.0,
+ 1728.0,
+ 1404.0,
+ 1728.0,
+ 1404.0,
+ 1763.0,
+ 570.0,
+ 1763.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1758.0,
+ 1243.0,
+ 1758.0,
+ 1243.0,
+ 1793.0,
+ 294.0,
+ 1793.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1264.0,
+ 1758.0,
+ 1403.0,
+ 1758.0,
+ 1403.0,
+ 1793.0,
+ 1264.0,
+ 1793.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1791.0,
+ 1403.0,
+ 1791.0,
+ 1403.0,
+ 1822.0,
+ 296.0,
+ 1822.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1818.0,
+ 1406.0,
+ 1818.0,
+ 1406.0,
+ 1856.0,
+ 292.0,
+ 1856.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1851.0,
+ 297.0,
+ 1851.0,
+ 297.0,
+ 1884.0,
+ 294.0,
+ 1884.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 481.0,
+ 1851.0,
+ 568.0,
+ 1851.0,
+ 568.0,
+ 1884.0,
+ 481.0,
+ 1884.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 589.0,
+ 1851.0,
+ 1406.0,
+ 1851.0,
+ 1406.0,
+ 1884.0,
+ 589.0,
+ 1884.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 285.0,
+ 1882.0,
+ 586.0,
+ 1882.0,
+ 586.0,
+ 1962.0,
+ 285.0,
+ 1962.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 767.0,
+ 1882.0,
+ 1147.0,
+ 1882.0,
+ 1147.0,
+ 1962.0,
+ 767.0,
+ 1962.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1318.0,
+ 1882.0,
+ 1416.0,
+ 1882.0,
+ 1416.0,
+ 1962.0,
+ 1318.0,
+ 1962.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1942.0,
+ 1406.0,
+ 1942.0,
+ 1406.0,
+ 1976.0,
+ 294.0,
+ 1976.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1972.0,
+ 482.0,
+ 1972.0,
+ 482.0,
+ 2009.0,
+ 291.0,
+ 2009.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 566.0,
+ 1972.0,
+ 575.0,
+ 1972.0,
+ 575.0,
+ 2009.0,
+ 566.0,
+ 2009.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 771.0,
+ 1972.0,
+ 854.0,
+ 1972.0,
+ 854.0,
+ 2009.0,
+ 771.0,
+ 2009.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 971.0,
+ 1972.0,
+ 1411.0,
+ 1972.0,
+ 1411.0,
+ 2009.0,
+ 971.0,
+ 2009.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 2001.0,
+ 430.0,
+ 2001.0,
+ 430.0,
+ 2039.0,
+ 294.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 459.0,
+ 2001.0,
+ 1133.0,
+ 2001.0,
+ 1133.0,
+ 2039.0,
+ 459.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1169.0,
+ 2001.0,
+ 1407.0,
+ 2001.0,
+ 1407.0,
+ 2039.0,
+ 1169.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1363.0,
+ 373.0,
+ 1363.0,
+ 373.0,
+ 1403.0,
+ 292.0,
+ 1403.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 458.0,
+ 1363.0,
+ 1406.0,
+ 1363.0,
+ 1406.0,
+ 1403.0,
+ 458.0,
+ 1403.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1396.0,
+ 1408.0,
+ 1396.0,
+ 1408.0,
+ 1430.0,
+ 294.0,
+ 1430.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1422.0,
+ 1405.0,
+ 1422.0,
+ 1405.0,
+ 1465.0,
+ 291.0,
+ 1465.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1456.0,
+ 1405.0,
+ 1456.0,
+ 1405.0,
+ 1491.0,
+ 294.0,
+ 1491.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1489.0,
+ 1403.0,
+ 1489.0,
+ 1403.0,
+ 1522.0,
+ 294.0,
+ 1522.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1516.0,
+ 1408.0,
+ 1516.0,
+ 1408.0,
+ 1555.0,
+ 291.0,
+ 1555.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 283.0,
+ 1536.0,
+ 981.0,
+ 1536.0,
+ 981.0,
+ 1613.0,
+ 283.0,
+ 1613.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1066.0,
+ 1536.0,
+ 1198.0,
+ 1536.0,
+ 1198.0,
+ 1613.0,
+ 1066.0,
+ 1613.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1297.0,
+ 1536.0,
+ 1415.0,
+ 1536.0,
+ 1415.0,
+ 1613.0,
+ 1297.0,
+ 1613.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 312.0,
+ 1404.0,
+ 312.0,
+ 1404.0,
+ 349.0,
+ 295.0,
+ 349.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 316.0,
+ 346.0,
+ 1027.0,
+ 346.0,
+ 1027.0,
+ 380.0,
+ 316.0,
+ 380.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 383.0,
+ 1406.0,
+ 383.0,
+ 1406.0,
+ 427.0,
+ 292.0,
+ 427.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 318.0,
+ 421.0,
+ 1340.0,
+ 421.0,
+ 1340.0,
+ 455.0,
+ 318.0,
+ 455.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 460.0,
+ 1406.0,
+ 460.0,
+ 1406.0,
+ 500.0,
+ 292.0,
+ 500.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 318.0,
+ 493.0,
+ 857.0,
+ 493.0,
+ 857.0,
+ 528.0,
+ 318.0,
+ 528.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 230.0,
+ 1405.0,
+ 230.0,
+ 1405.0,
+ 266.0,
+ 295.0,
+ 266.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 262.0,
+ 607.0,
+ 262.0,
+ 607.0,
+ 294.0,
+ 295.0,
+ 294.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1260.0,
+ 969.0,
+ 1260.0,
+ 969.0,
+ 1299.0,
+ 294.0,
+ 1299.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 1,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 714,
+ 826,
+ 714,
+ 826,
+ 1324,
+ 299,
+ 1324
+ ],
+ "score": 0.98
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 399,
+ 1405,
+ 399,
+ 1405,
+ 621,
+ 297,
+ 621
+ ],
+ "score": 0.98
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 1419,
+ 825,
+ 1419,
+ 825,
+ 1874,
+ 299,
+ 1874
+ ],
+ "score": 0.976
+ },
+ {
+ "category_id": 3,
+ "poly": [
+ 851,
+ 645,
+ 1402,
+ 645,
+ 1402,
+ 1459,
+ 851,
+ 1459
+ ],
+ "score": 0.971
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 300,
+ 1943,
+ 1400,
+ 1943,
+ 1400,
+ 2034,
+ 300,
+ 2034
+ ],
+ "score": 0.963
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 848,
+ 1476,
+ 1404,
+ 1476,
+ 1404,
+ 1843,
+ 848,
+ 1843
+ ],
+ "score": 0.96
+ },
+ {
+ "category_id": 8,
+ "poly": [
+ 635,
+ 298,
+ 1062,
+ 298,
+ 1062,
+ 385,
+ 635,
+ 385
+ ],
+ "score": 0.957
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 295,
+ 228,
+ 1403,
+ 228,
+ 1403,
+ 292,
+ 295,
+ 292
+ ],
+ "score": 0.945
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 299,
+ 74,
+ 813,
+ 74,
+ 813,
+ 105,
+ 299,
+ 105
+ ],
+ "score": 0.913
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 302,
+ 659,
+ 785,
+ 659,
+ 785,
+ 689,
+ 302,
+ 689
+ ],
+ "score": 0.895
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 302,
+ 1875,
+ 1200,
+ 1875,
+ 1200,
+ 1906,
+ 302,
+ 1906
+ ],
+ "score": 0.892
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 303,
+ 1365,
+ 751,
+ 1365,
+ 751,
+ 1393,
+ 303,
+ 1393
+ ],
+ "score": 0.887
+ },
+ {
+ "category_id": 9,
+ "poly": [
+ 1366,
+ 323,
+ 1400,
+ 323,
+ 1400,
+ 352,
+ 1366,
+ 352
+ ],
+ "score": 0.869
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 858,
+ 2088,
+ 858,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.623
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 859,
+ 2088,
+ 859,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.421
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 848,
+ 1477,
+ 1403,
+ 1477,
+ 1403,
+ 1843,
+ 848,
+ 1843
+ ],
+ "score": 0.163
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 635,
+ 295,
+ 1065,
+ 295,
+ 1065,
+ 384,
+ 635,
+ 384
+ ],
+ "score": 0.95,
+ "latex": "w _ { t } ^ { n } = \\frac { \\exp { \\left( - \\sigma ^ { - 2 } ( t - c _ { n } ) ^ { 2 } \\right) } } { \\sum _ { m = 1 } ^ { N } \\exp { \\left( - \\sigma ^ { - 2 } ( t - c _ { m } ) ^ { 2 } \\right) } } ."
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 806,
+ 399,
+ 1023,
+ 399,
+ 1023,
+ 442,
+ 806,
+ 442
+ ],
+ "score": 0.93,
+ "latex": "\\begin{array} { r } { a _ { t } \\ = \\ \\sum _ { n = 1 } ^ { N } w _ { t } ^ { n } h _ { n } } \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 883,
+ 1752,
+ 952,
+ 1752,
+ 952,
+ 1783,
+ 883,
+ 1783
+ ],
+ "score": 0.91,
+ "latex": "\\eta = 0"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1125,
+ 229,
+ 1156,
+ 229,
+ 1156,
+ 258,
+ 1125,
+ 258
+ ],
+ "score": 0.87,
+ "latex": "\\sigma ^ { 2 }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 649,
+ 1172,
+ 676,
+ 1172,
+ 676,
+ 1201,
+ 649,
+ 1201
+ ],
+ "score": 0.87,
+ "latex": "l _ { n }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 694,
+ 236,
+ 724,
+ 236,
+ 724,
+ 261,
+ 694,
+ 261
+ ],
+ "score": 0.86,
+ "latex": "c _ { n }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 433,
+ 1114,
+ 462,
+ 1114,
+ 462,
+ 1140,
+ 433,
+ 1140
+ ],
+ "score": 0.85,
+ "latex": "a _ { t }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 937,
+ 529,
+ 956,
+ 529,
+ 956,
+ 560,
+ 937,
+ 560
+ ],
+ "score": 0.84,
+ "latex": "f"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1020,
+ 1665,
+ 1040,
+ 1665,
+ 1040,
+ 1692,
+ 1020,
+ 1692
+ ],
+ "score": 0.8,
+ "latex": "\\eta"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 630,
+ 235,
+ 644,
+ 235,
+ 644,
+ 258,
+ 630,
+ 258
+ ],
+ "score": 0.76,
+ "latex": "t"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 428,
+ 1662,
+ 519,
+ 1662,
+ 519,
+ 1691,
+ 428,
+ 1691
+ ],
+ "score": 0.74,
+ "latex": "2 4 ~ \\mathrm { k H z }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1297,
+ 1513,
+ 1317,
+ 1513,
+ 1317,
+ 1536,
+ 1297,
+ 1536
+ ],
+ "score": 0.74,
+ "latex": "z"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 664,
+ 994,
+ 683,
+ 994,
+ 683,
+ 1021,
+ 664,
+ 1021
+ ],
+ "score": 0.72,
+ "latex": "\\eta"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 422,
+ 1019,
+ 506,
+ 1019,
+ 506,
+ 1048,
+ 422,
+ 1048
+ ],
+ "score": 0.72,
+ "latex": "2 0 0 \\mathrm { H z }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 631,
+ 1112,
+ 647,
+ 1112,
+ 647,
+ 1137,
+ 631,
+ 1137
+ ],
+ "score": 0.68,
+ "latex": "t"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 609,
+ 836,
+ 695,
+ 836,
+ 695,
+ 865,
+ 609,
+ 865
+ ],
+ "score": 0.67,
+ "latex": "2 4 \\mathrm { k H z }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 702,
+ 1662,
+ 791,
+ 1662,
+ 791,
+ 1692,
+ 702,
+ 1692
+ ],
+ "score": 0.48,
+ "latex": "1 5 0 ~ \\mathrm { m s }"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 873.0,
+ 681.0,
+ 1031.0,
+ 681.0,
+ 1031.0,
+ 734.0,
+ 873.0,
+ 734.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1114.0,
+ 693.0,
+ 1242.0,
+ 693.0,
+ 1242.0,
+ 732.0,
+ 1114.0,
+ 732.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 869.0,
+ 732.0,
+ 1062.0,
+ 732.0,
+ 1062.0,
+ 770.0,
+ 869.0,
+ 770.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1110.0,
+ 723.0,
+ 1248.0,
+ 723.0,
+ 1248.0,
+ 765.0,
+ 1110.0,
+ 765.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 896.0,
+ 776.0,
+ 997.0,
+ 776.0,
+ 997.0,
+ 819.0,
+ 896.0,
+ 819.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1118.0,
+ 780.0,
+ 1238.0,
+ 780.0,
+ 1238.0,
+ 816.0,
+ 1118.0,
+ 816.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1229.0,
+ 816.0,
+ 1365.0,
+ 816.0,
+ 1365.0,
+ 851.0,
+ 1229.0,
+ 851.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1031.0,
+ 850.0,
+ 1325.0,
+ 850.0,
+ 1325.0,
+ 890.0,
+ 1031.0,
+ 890.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1116.0,
+ 902.0,
+ 1275.0,
+ 902.0,
+ 1275.0,
+ 938.0,
+ 1116.0,
+ 938.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1116.0,
+ 929.0,
+ 1279.0,
+ 929.0,
+ 1279.0,
+ 970.0,
+ 1116.0,
+ 970.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 920.0,
+ 934.0,
+ 971.0,
+ 934.0,
+ 971.0,
+ 1070.0,
+ 920.0,
+ 1070.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1031.0,
+ 974.0,
+ 1101.0,
+ 974.0,
+ 1101.0,
+ 997.0,
+ 1031.0,
+ 997.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1028.0,
+ 989.0,
+ 1099.0,
+ 989.0,
+ 1099.0,
+ 1013.0,
+ 1028.0,
+ 1013.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1035.0,
+ 1019.0,
+ 1077.0,
+ 1019.0,
+ 1077.0,
+ 1045.0,
+ 1035.0,
+ 1045.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1283.0,
+ 1021.0,
+ 1343.0,
+ 1021.0,
+ 1343.0,
+ 1041.0,
+ 1283.0,
+ 1041.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1298.0,
+ 1048.0,
+ 1318.0,
+ 1048.0,
+ 1318.0,
+ 1064.0,
+ 1298.0,
+ 1064.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1007.0,
+ 1082.0,
+ 1071.0,
+ 1082.0,
+ 1071.0,
+ 1112.0,
+ 1007.0,
+ 1112.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1104.0,
+ 1078.0,
+ 1252.0,
+ 1078.0,
+ 1252.0,
+ 1112.0,
+ 1104.0,
+ 1112.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1281.0,
+ 1080.0,
+ 1349.0,
+ 1080.0,
+ 1349.0,
+ 1116.0,
+ 1281.0,
+ 1116.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1080.0,
+ 1108.0,
+ 1270.0,
+ 1108.0,
+ 1270.0,
+ 1152.0,
+ 1080.0,
+ 1152.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1005.0,
+ 1148.0,
+ 1073.0,
+ 1148.0,
+ 1073.0,
+ 1185.0,
+ 1005.0,
+ 1185.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1105.0,
+ 1148.0,
+ 1251.0,
+ 1148.0,
+ 1251.0,
+ 1182.0,
+ 1105.0,
+ 1182.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1279.0,
+ 1150.0,
+ 1348.0,
+ 1150.0,
+ 1348.0,
+ 1184.0,
+ 1279.0,
+ 1184.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 898.0,
+ 1180.0,
+ 995.0,
+ 1180.0,
+ 995.0,
+ 1219.0,
+ 898.0,
+ 1219.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1061.0,
+ 1181.0,
+ 1299.0,
+ 1181.0,
+ 1299.0,
+ 1219.0,
+ 1061.0,
+ 1219.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 897.0,
+ 1213.0,
+ 996.0,
+ 1213.0,
+ 996.0,
+ 1256.0,
+ 897.0,
+ 1256.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1053.0,
+ 1215.0,
+ 1306.0,
+ 1215.0,
+ 1306.0,
+ 1254.0,
+ 1053.0,
+ 1254.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 881.0,
+ 1275.0,
+ 984.0,
+ 1275.0,
+ 984.0,
+ 1311.0,
+ 881.0,
+ 1311.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1007.0,
+ 1250.0,
+ 1371.0,
+ 1250.0,
+ 1371.0,
+ 1312.0,
+ 1007.0,
+ 1312.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 896.0,
+ 1305.0,
+ 999.0,
+ 1305.0,
+ 999.0,
+ 1348.0,
+ 896.0,
+ 1348.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1111.0,
+ 1309.0,
+ 1246.0,
+ 1309.0,
+ 1246.0,
+ 1348.0,
+ 1111.0,
+ 1348.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1004.0,
+ 1355.0,
+ 1335.0,
+ 1355.0,
+ 1335.0,
+ 1391.0,
+ 1004.0,
+ 1391.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 875.0,
+ 1396.0,
+ 904.0,
+ 1396.0,
+ 904.0,
+ 1425.0,
+ 875.0,
+ 1425.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1100.0,
+ 1390.0,
+ 1257.0,
+ 1390.0,
+ 1257.0,
+ 1427.0,
+ 1100.0,
+ 1427.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 871.0,
+ 1428.0,
+ 995.0,
+ 1428.0,
+ 995.0,
+ 1461.0,
+ 871.0,
+ 1461.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1007.0,
+ 1427.0,
+ 1346.0,
+ 1427.0,
+ 1346.0,
+ 1462.0,
+ 1007.0,
+ 1462.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1367.0,
+ 1431.0,
+ 1395.0,
+ 1431.0,
+ 1395.0,
+ 1460.0,
+ 1367.0,
+ 1460.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1086.0,
+ 943.5,
+ 1095.0,
+ 943.5,
+ 1095.0,
+ 966.0,
+ 1086.0,
+ 966.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 845.0,
+ 1476.0,
+ 1403.0,
+ 1476.0,
+ 1403.0,
+ 1510.0,
+ 845.0,
+ 1510.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 845.0,
+ 1507.0,
+ 1296.0,
+ 1507.0,
+ 1296.0,
+ 1540.0,
+ 845.0,
+ 1540.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1318.0,
+ 1507.0,
+ 1404.0,
+ 1507.0,
+ 1404.0,
+ 1540.0,
+ 1318.0,
+ 1540.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 845.0,
+ 1537.0,
+ 1404.0,
+ 1537.0,
+ 1404.0,
+ 1569.0,
+ 845.0,
+ 1569.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 844.0,
+ 1567.0,
+ 1407.0,
+ 1567.0,
+ 1407.0,
+ 1603.0,
+ 844.0,
+ 1603.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 845.0,
+ 1599.0,
+ 1404.0,
+ 1599.0,
+ 1404.0,
+ 1629.0,
+ 845.0,
+ 1629.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 843.0,
+ 1625.0,
+ 1406.0,
+ 1625.0,
+ 1406.0,
+ 1664.0,
+ 843.0,
+ 1664.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 843.0,
+ 1658.0,
+ 1019.0,
+ 1658.0,
+ 1019.0,
+ 1694.0,
+ 843.0,
+ 1694.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1041.0,
+ 1658.0,
+ 1407.0,
+ 1658.0,
+ 1407.0,
+ 1694.0,
+ 1041.0,
+ 1694.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 844.0,
+ 1689.0,
+ 1405.0,
+ 1689.0,
+ 1405.0,
+ 1721.0,
+ 844.0,
+ 1721.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 844.0,
+ 1721.0,
+ 1406.0,
+ 1721.0,
+ 1406.0,
+ 1754.0,
+ 844.0,
+ 1754.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 843.0,
+ 1750.0,
+ 882.0,
+ 1750.0,
+ 882.0,
+ 1785.0,
+ 843.0,
+ 1785.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 953.0,
+ 1750.0,
+ 1408.0,
+ 1750.0,
+ 1408.0,
+ 1785.0,
+ 953.0,
+ 1785.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 844.0,
+ 1782.0,
+ 1405.0,
+ 1782.0,
+ 1405.0,
+ 1814.0,
+ 844.0,
+ 1814.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 845.0,
+ 1809.0,
+ 1119.0,
+ 1809.0,
+ 1119.0,
+ 1846.0,
+ 845.0,
+ 1846.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 72.0,
+ 815.0,
+ 72.0,
+ 815.0,
+ 108.0,
+ 295.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 658.0,
+ 788.0,
+ 658.0,
+ 788.0,
+ 693.0,
+ 296.0,
+ 693.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1363.0,
+ 756.0,
+ 1363.0,
+ 756.0,
+ 1397.0,
+ 296.0,
+ 1397.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 839.0,
+ 2085.0,
+ 860.0,
+ 2085.0,
+ 860.0,
+ 2116.0,
+ 839.0,
+ 2116.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 838.0,
+ 2085.0,
+ 862.0,
+ 2085.0,
+ 862.0,
+ 2118.0,
+ 838.0,
+ 2118.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 845.0,
+ 1477.0,
+ 1403.0,
+ 1477.0,
+ 1403.0,
+ 1510.0,
+ 845.0,
+ 1510.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 845.0,
+ 1508.0,
+ 1296.0,
+ 1508.0,
+ 1296.0,
+ 1541.0,
+ 845.0,
+ 1541.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1318.0,
+ 1508.0,
+ 1404.0,
+ 1508.0,
+ 1404.0,
+ 1541.0,
+ 1318.0,
+ 1541.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 844.0,
+ 1536.0,
+ 1405.0,
+ 1536.0,
+ 1405.0,
+ 1571.0,
+ 844.0,
+ 1571.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 844.0,
+ 1566.0,
+ 1407.0,
+ 1566.0,
+ 1407.0,
+ 1605.0,
+ 844.0,
+ 1605.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 846.0,
+ 1600.0,
+ 1404.0,
+ 1600.0,
+ 1404.0,
+ 1630.0,
+ 846.0,
+ 1630.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 843.0,
+ 1626.0,
+ 1406.0,
+ 1626.0,
+ 1406.0,
+ 1665.0,
+ 843.0,
+ 1665.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 843.0,
+ 1658.0,
+ 1019.0,
+ 1658.0,
+ 1019.0,
+ 1693.0,
+ 843.0,
+ 1693.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1041.0,
+ 1658.0,
+ 1406.0,
+ 1658.0,
+ 1406.0,
+ 1693.0,
+ 1041.0,
+ 1693.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 844.0,
+ 1689.0,
+ 1404.0,
+ 1689.0,
+ 1404.0,
+ 1721.0,
+ 844.0,
+ 1721.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 843.0,
+ 1720.0,
+ 1405.0,
+ 1720.0,
+ 1405.0,
+ 1755.0,
+ 843.0,
+ 1755.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 844.0,
+ 1751.0,
+ 882.0,
+ 1751.0,
+ 882.0,
+ 1784.0,
+ 844.0,
+ 1784.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 953.0,
+ 1751.0,
+ 1408.0,
+ 1751.0,
+ 1408.0,
+ 1784.0,
+ 953.0,
+ 1784.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 843.0,
+ 1782.0,
+ 1406.0,
+ 1782.0,
+ 1406.0,
+ 1813.0,
+ 843.0,
+ 1813.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 845.0,
+ 1809.0,
+ 1118.0,
+ 1809.0,
+ 1118.0,
+ 1847.0,
+ 845.0,
+ 1847.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 714.0,
+ 826.0,
+ 714.0,
+ 826.0,
+ 749.0,
+ 296.0,
+ 749.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 743.0,
+ 828.0,
+ 743.0,
+ 828.0,
+ 779.0,
+ 295.0,
+ 779.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 775.0,
+ 827.0,
+ 775.0,
+ 827.0,
+ 810.0,
+ 295.0,
+ 810.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 803.0,
+ 830.0,
+ 803.0,
+ 830.0,
+ 841.0,
+ 294.0,
+ 841.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 835.0,
+ 608.0,
+ 835.0,
+ 608.0,
+ 870.0,
+ 295.0,
+ 870.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 696.0,
+ 835.0,
+ 827.0,
+ 835.0,
+ 827.0,
+ 870.0,
+ 696.0,
+ 870.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 866.0,
+ 829.0,
+ 866.0,
+ 829.0,
+ 899.0,
+ 295.0,
+ 899.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 898.0,
+ 827.0,
+ 898.0,
+ 827.0,
+ 929.0,
+ 295.0,
+ 929.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 927.0,
+ 825.0,
+ 927.0,
+ 825.0,
+ 958.0,
+ 295.0,
+ 958.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 958.0,
+ 829.0,
+ 958.0,
+ 829.0,
+ 990.0,
+ 295.0,
+ 990.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 989.0,
+ 663.0,
+ 989.0,
+ 663.0,
+ 1023.0,
+ 295.0,
+ 1023.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 684.0,
+ 989.0,
+ 829.0,
+ 989.0,
+ 829.0,
+ 1023.0,
+ 684.0,
+ 1023.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1020.0,
+ 421.0,
+ 1020.0,
+ 421.0,
+ 1051.0,
+ 296.0,
+ 1051.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 507.0,
+ 1020.0,
+ 827.0,
+ 1020.0,
+ 827.0,
+ 1051.0,
+ 507.0,
+ 1051.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1049.0,
+ 828.0,
+ 1049.0,
+ 828.0,
+ 1081.0,
+ 296.0,
+ 1081.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1081.0,
+ 827.0,
+ 1081.0,
+ 827.0,
+ 1112.0,
+ 296.0,
+ 1112.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1112.0,
+ 432.0,
+ 1112.0,
+ 432.0,
+ 1141.0,
+ 295.0,
+ 1141.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 463.0,
+ 1112.0,
+ 630.0,
+ 1112.0,
+ 630.0,
+ 1141.0,
+ 463.0,
+ 1141.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 648.0,
+ 1112.0,
+ 826.0,
+ 1112.0,
+ 826.0,
+ 1141.0,
+ 648.0,
+ 1141.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1141.0,
+ 829.0,
+ 1141.0,
+ 829.0,
+ 1173.0,
+ 295.0,
+ 1173.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1171.0,
+ 648.0,
+ 1171.0,
+ 648.0,
+ 1204.0,
+ 295.0,
+ 1204.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 677.0,
+ 1171.0,
+ 825.0,
+ 1171.0,
+ 825.0,
+ 1204.0,
+ 677.0,
+ 1204.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1204.0,
+ 825.0,
+ 1204.0,
+ 825.0,
+ 1235.0,
+ 295.0,
+ 1235.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1233.0,
+ 825.0,
+ 1233.0,
+ 825.0,
+ 1264.0,
+ 296.0,
+ 1264.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1263.0,
+ 825.0,
+ 1263.0,
+ 825.0,
+ 1294.0,
+ 296.0,
+ 1294.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1292.0,
+ 785.0,
+ 1292.0,
+ 785.0,
+ 1327.0,
+ 294.0,
+ 1327.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 285.0,
+ 390.0,
+ 805.0,
+ 390.0,
+ 805.0,
+ 460.0,
+ 285.0,
+ 460.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1024.0,
+ 390.0,
+ 1414.0,
+ 390.0,
+ 1414.0,
+ 460.0,
+ 1024.0,
+ 460.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 436.0,
+ 1408.0,
+ 436.0,
+ 1408.0,
+ 475.0,
+ 294.0,
+ 475.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 468.0,
+ 1403.0,
+ 468.0,
+ 1403.0,
+ 500.0,
+ 296.0,
+ 500.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 497.0,
+ 1403.0,
+ 497.0,
+ 1403.0,
+ 533.0,
+ 295.0,
+ 533.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 528.0,
+ 936.0,
+ 528.0,
+ 936.0,
+ 563.0,
+ 295.0,
+ 563.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 957.0,
+ 528.0,
+ 1405.0,
+ 528.0,
+ 1405.0,
+ 563.0,
+ 957.0,
+ 563.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 560.0,
+ 1403.0,
+ 560.0,
+ 1403.0,
+ 592.0,
+ 296.0,
+ 592.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 588.0,
+ 1388.0,
+ 588.0,
+ 1388.0,
+ 626.0,
+ 292.0,
+ 626.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1417.0,
+ 826.0,
+ 1417.0,
+ 826.0,
+ 1451.0,
+ 296.0,
+ 1451.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1449.0,
+ 827.0,
+ 1449.0,
+ 827.0,
+ 1480.0,
+ 296.0,
+ 1480.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1478.0,
+ 826.0,
+ 1478.0,
+ 826.0,
+ 1510.0,
+ 296.0,
+ 1510.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1508.0,
+ 830.0,
+ 1508.0,
+ 830.0,
+ 1545.0,
+ 293.0,
+ 1545.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1540.0,
+ 831.0,
+ 1540.0,
+ 831.0,
+ 1572.0,
+ 295.0,
+ 1572.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1569.0,
+ 831.0,
+ 1569.0,
+ 831.0,
+ 1603.0,
+ 294.0,
+ 1603.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1602.0,
+ 827.0,
+ 1602.0,
+ 827.0,
+ 1630.0,
+ 296.0,
+ 1630.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1632.0,
+ 828.0,
+ 1632.0,
+ 828.0,
+ 1662.0,
+ 295.0,
+ 1662.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1663.0,
+ 427.0,
+ 1663.0,
+ 427.0,
+ 1692.0,
+ 294.0,
+ 1692.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 520.0,
+ 1663.0,
+ 701.0,
+ 1663.0,
+ 701.0,
+ 1692.0,
+ 520.0,
+ 1692.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 792.0,
+ 1663.0,
+ 828.0,
+ 1663.0,
+ 828.0,
+ 1692.0,
+ 792.0,
+ 1692.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1692.0,
+ 829.0,
+ 1692.0,
+ 829.0,
+ 1724.0,
+ 295.0,
+ 1724.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1722.0,
+ 827.0,
+ 1722.0,
+ 827.0,
+ 1754.0,
+ 295.0,
+ 1754.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1755.0,
+ 829.0,
+ 1755.0,
+ 829.0,
+ 1784.0,
+ 295.0,
+ 1784.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1785.0,
+ 827.0,
+ 1785.0,
+ 827.0,
+ 1816.0,
+ 295.0,
+ 1816.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1813.0,
+ 825.0,
+ 1813.0,
+ 825.0,
+ 1847.0,
+ 295.0,
+ 1847.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1841.0,
+ 830.0,
+ 1841.0,
+ 830.0,
+ 1881.0,
+ 294.0,
+ 1881.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1938.0,
+ 1406.0,
+ 1938.0,
+ 1406.0,
+ 1981.0,
+ 293.0,
+ 1981.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1972.0,
+ 1404.0,
+ 1972.0,
+ 1404.0,
+ 2009.0,
+ 294.0,
+ 2009.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 2002.0,
+ 1405.0,
+ 2002.0,
+ 1405.0,
+ 2039.0,
+ 295.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 227.0,
+ 629.0,
+ 227.0,
+ 629.0,
+ 265.0,
+ 293.0,
+ 265.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 645.0,
+ 227.0,
+ 693.0,
+ 227.0,
+ 693.0,
+ 265.0,
+ 645.0,
+ 265.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 725.0,
+ 227.0,
+ 1124.0,
+ 227.0,
+ 1124.0,
+ 265.0,
+ 725.0,
+ 265.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1157.0,
+ 227.0,
+ 1406.0,
+ 227.0,
+ 1406.0,
+ 265.0,
+ 1157.0,
+ 265.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 260.0,
+ 568.0,
+ 260.0,
+ 568.0,
+ 296.0,
+ 294.0,
+ 296.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 1872.0,
+ 1205.0,
+ 1872.0,
+ 1205.0,
+ 1911.0,
+ 297.0,
+ 1911.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 2,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 376,
+ 1405,
+ 376,
+ 1405,
+ 652,
+ 297,
+ 652
+ ],
+ "score": 0.983
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 661,
+ 1403,
+ 661,
+ 1403,
+ 847,
+ 297,
+ 847
+ ],
+ "score": 0.981
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 941,
+ 1404,
+ 941,
+ 1404,
+ 1159,
+ 297,
+ 1159
+ ],
+ "score": 0.98
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1167,
+ 1404,
+ 1167,
+ 1404,
+ 1353,
+ 298,
+ 1353
+ ],
+ "score": 0.979
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1362,
+ 1403,
+ 1362,
+ 1403,
+ 1516,
+ 298,
+ 1516
+ ],
+ "score": 0.978
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1599,
+ 1405,
+ 1599,
+ 1405,
+ 1785,
+ 298,
+ 1785
+ ],
+ "score": 0.977
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 302,
+ 1941,
+ 1399,
+ 1941,
+ 1399,
+ 2035,
+ 302,
+ 2035
+ ],
+ "score": 0.961
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 294,
+ 229,
+ 1404,
+ 229,
+ 1404,
+ 293,
+ 294,
+ 293
+ ],
+ "score": 0.951
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 363,
+ 1804,
+ 1370,
+ 1804,
+ 1370,
+ 1926,
+ 363,
+ 1926
+ ],
+ "score": 0.934
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 300,
+ 75,
+ 813,
+ 75,
+ 813,
+ 105,
+ 300,
+ 105
+ ],
+ "score": 0.903
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 299,
+ 1548,
+ 669,
+ 1548,
+ 669,
+ 1579,
+ 299,
+ 1579
+ ],
+ "score": 0.902
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 302,
+ 325,
+ 765,
+ 325,
+ 765,
+ 356,
+ 302,
+ 356
+ ],
+ "score": 0.896
+ },
+ {
+ "category_id": 9,
+ "poly": [
+ 1366,
+ 888,
+ 1399,
+ 888,
+ 1399,
+ 918,
+ 1366,
+ 918
+ ],
+ "score": 0.888
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 840,
+ 2088,
+ 857,
+ 2088,
+ 857,
+ 2111,
+ 840,
+ 2111
+ ],
+ "score": 0.78
+ },
+ {
+ "category_id": 8,
+ "poly": [
+ 581,
+ 878,
+ 1118,
+ 878,
+ 1118,
+ 928,
+ 581,
+ 928
+ ],
+ "score": 0.661
+ },
+ {
+ "category_id": 8,
+ "poly": [
+ 580,
+ 878,
+ 1119,
+ 878,
+ 1119,
+ 928,
+ 580,
+ 928
+ ],
+ "score": 0.369
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1039,
+ 785,
+ 1114,
+ 785,
+ 1114,
+ 818,
+ 1039,
+ 818
+ ],
+ "score": 0.93,
+ "latex": "S [ t , f ]"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1189,
+ 1725,
+ 1399,
+ 1725,
+ 1399,
+ 1757,
+ 1189,
+ 1757
+ ],
+ "score": 0.92,
+ "latex": "p _ { \\mathrm { g e n , 1 } } = p _ { \\mathrm { g t , 1 } } = 1"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1153,
+ 1973,
+ 1262,
+ 1973,
+ 1262,
+ 2007,
+ 1153,
+ 2007
+ ],
+ "score": 0.92,
+ "latex": "\\mathrm { \\dot { \\cal S } } _ { \\mathrm { g t } } [ p _ { \\mathrm { g t } , k } ]"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 581,
+ 877,
+ 1120,
+ 877,
+ 1120,
+ 927,
+ 581,
+ 927
+ ],
+ "score": 0.92,
+ "latex": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { p r e d } } = \\frac { 1 } { F } \\sum _ { t = 1 } ^ { T } \\sum _ { f = 1 } ^ { F } | S _ { \\mathrm { g e n } } [ t , f ] - S _ { \\mathrm { g t } } [ t , f ] | . } \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 964,
+ 1974,
+ 1101,
+ 1974,
+ 1101,
+ 2007,
+ 964,
+ 2007
+ ],
+ "score": 0.91,
+ "latex": "S _ { \\mathrm { g e n } } [ p _ { \\mathrm { g e n } , k } ]"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 932,
+ 755,
+ 986,
+ 755,
+ 986,
+ 788,
+ 932,
+ 788
+ ],
+ "score": 0.91,
+ "latex": "S _ { \\mathrm { g e n } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1039,
+ 1363,
+ 1106,
+ 1363,
+ 1106,
+ 1395,
+ 1039,
+ 1395
+ ],
+ "score": 0.91,
+ "latex": "\\mathcal { L } _ { \\mathrm { p r e d } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 692,
+ 1890,
+ 745,
+ 1890,
+ 745,
+ 1923,
+ 692,
+ 1923
+ ],
+ "score": 0.91,
+ "latex": "S _ { \\mathrm { g e n } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 459,
+ 1723,
+ 513,
+ 1723,
+ 513,
+ 1755,
+ 459,
+ 1755
+ ],
+ "score": 0.9,
+ "latex": "S _ { \\mathrm { g e n } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 500,
+ 1127,
+ 541,
+ 1127,
+ 541,
+ 1161,
+ 500,
+ 1161
+ ],
+ "score": 0.9,
+ "latex": "S _ { \\mathrm { g t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 564,
+ 1723,
+ 605,
+ 1723,
+ 605,
+ 1756,
+ 564,
+ 1756
+ ],
+ "score": 0.9,
+ "latex": "S _ { \\mathrm { g t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 373,
+ 785,
+ 414,
+ 785,
+ 414,
+ 818,
+ 373,
+ 818
+ ],
+ "score": 0.9,
+ "latex": "S _ { \\mathrm { g t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 546,
+ 1942,
+ 1035,
+ 1942,
+ 1035,
+ 1978,
+ 546,
+ 1978
+ ],
+ "score": 0.89,
+ "latex": "p = \\langle ( p _ { \\mathrm { g e n } , 1 } , p _ { \\mathrm { g t } , 1 } ) , \\dots , ( p _ { \\mathrm { g e n } , K _ { p } } , p _ { \\mathrm { g t } , K _ { p } } ) \\rangle"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 691,
+ 1848,
+ 732,
+ 1848,
+ 732,
+ 1882,
+ 691,
+ 1882
+ ],
+ "score": 0.89,
+ "latex": "S _ { \\mathrm { g t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 488,
+ 695,
+ 522,
+ 695,
+ 522,
+ 724,
+ 488,
+ 724
+ ],
+ "score": 0.89,
+ "latex": "L _ { 1 }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1128,
+ 1944,
+ 1167,
+ 1944,
+ 1167,
+ 1975,
+ 1128,
+ 1975
+ ],
+ "score": 0.89,
+ "latex": "K _ { p }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 724,
+ 1976,
+ 758,
+ 1976,
+ 758,
+ 2003,
+ 724,
+ 2003
+ ],
+ "score": 0.86,
+ "latex": "L _ { 1 }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 776,
+ 1097,
+ 830,
+ 1097,
+ 830,
+ 1126,
+ 776,
+ 1126
+ ],
+ "score": 0.86,
+ "latex": "\\pm 6 0"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 375,
+ 945,
+ 400,
+ 945,
+ 400,
+ 971,
+ 375,
+ 971
+ ],
+ "score": 0.83,
+ "latex": "F"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 706,
+ 816,
+ 726,
+ 816,
+ 726,
+ 846,
+ 706,
+ 846
+ ],
+ "score": 0.83,
+ "latex": "f"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1025,
+ 1696,
+ 1043,
+ 1696,
+ 1043,
+ 1723,
+ 1025,
+ 1723
+ ],
+ "score": 0.82,
+ "latex": "p"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 488,
+ 1756,
+ 506,
+ 1756,
+ 506,
+ 1780,
+ 488,
+ 1780
+ ],
+ "score": 0.81,
+ "latex": "k"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 747,
+ 1806,
+ 800,
+ 1806,
+ 800,
+ 1840,
+ 747,
+ 1840
+ ],
+ "score": 0.81,
+ "latex": "S _ { \\mathrm { g e n } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 297,
+ 946,
+ 321,
+ 946,
+ 321,
+ 971,
+ 297,
+ 971
+ ],
+ "score": 0.8,
+ "latex": "T"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 815,
+ 1894,
+ 1065,
+ 1894,
+ 1065,
+ 1924,
+ 815,
+ 1924
+ ],
+ "score": 0.77,
+ "latex": "p _ { \\mathrm { g e n } , k + 1 } = p _ { \\mathrm { g e n } , k } + 1"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 432,
+ 818,
+ 445,
+ 818,
+ 445,
+ 842,
+ 432,
+ 842
+ ],
+ "score": 0.74,
+ "latex": "t"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 960,
+ 1097,
+ 1051,
+ 1097,
+ 1051,
+ 1126,
+ 960,
+ 1126
+ ],
+ "score": 0.72,
+ "latex": "2 4 \\mathrm { k H z }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 387,
+ 2008,
+ 410,
+ 2008,
+ 410,
+ 2030,
+ 387,
+ 2030
+ ],
+ "score": 0.71,
+ "latex": "w"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 812,
+ 1806,
+ 853,
+ 1806,
+ 853,
+ 1840,
+ 812,
+ 1840
+ ],
+ "score": 0.58,
+ "latex": "S _ { \\mathrm { g t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 798,
+ 1849,
+ 1245,
+ 1849,
+ 1245,
+ 1883,
+ 798,
+ 1883
+ ],
+ "score": 0.52,
+ "latex": "p _ { \\mathrm { g e n } , k + 1 } = p _ { \\mathrm { g e n } , k } , p _ { \\mathrm { g t } , k + 1 } = p _ { \\mathrm { g t } , k } + 1"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 802,
+ 1854,
+ 1004,
+ 1854,
+ 1004,
+ 1882,
+ 802,
+ 1882
+ ],
+ "score": 0.5,
+ "latex": "p _ { \\mathrm { g e n } , k + 1 } = p _ { \\mathrm { g e n } , k }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 741,
+ 1805,
+ 1364,
+ 1805,
+ 1364,
+ 1841,
+ 741,
+ 1841
+ ],
+ "score": 0.41,
+ "latex": "S _ { \\mathrm { g e n } } , S _ { \\mathrm { g t } } \\colon p _ { \\mathrm { g e n } , k + 1 } = p _ { \\mathrm { g e n } , k } + 1 , p _ { \\mathrm { g t } , k + 1 } = p _ { \\mathrm { g t } , k } + 1 ;"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 73.0,
+ 815.0,
+ 73.0,
+ 815.0,
+ 108.0,
+ 295.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1547.0,
+ 672.0,
+ 1547.0,
+ 672.0,
+ 1583.0,
+ 295.0,
+ 1583.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 324.0,
+ 770.0,
+ 324.0,
+ 770.0,
+ 360.0,
+ 296.0,
+ 360.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 838.0,
+ 2086.0,
+ 861.0,
+ 2086.0,
+ 861.0,
+ 2118.0,
+ 838.0,
+ 2118.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 378.0,
+ 1407.0,
+ 378.0,
+ 1407.0,
+ 411.0,
+ 296.0,
+ 411.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 408.0,
+ 1405.0,
+ 408.0,
+ 1405.0,
+ 441.0,
+ 295.0,
+ 441.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 440.0,
+ 1406.0,
+ 440.0,
+ 1406.0,
+ 472.0,
+ 296.0,
+ 472.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 470.0,
+ 1405.0,
+ 470.0,
+ 1405.0,
+ 502.0,
+ 296.0,
+ 502.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 501.0,
+ 1405.0,
+ 501.0,
+ 1405.0,
+ 534.0,
+ 295.0,
+ 534.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 528.0,
+ 1403.0,
+ 528.0,
+ 1403.0,
+ 564.0,
+ 294.0,
+ 564.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 561.0,
+ 1403.0,
+ 561.0,
+ 1403.0,
+ 594.0,
+ 295.0,
+ 594.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 588.0,
+ 1408.0,
+ 588.0,
+ 1408.0,
+ 626.0,
+ 292.0,
+ 626.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 622.0,
+ 753.0,
+ 622.0,
+ 753.0,
+ 655.0,
+ 296.0,
+ 655.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 662.0,
+ 1407.0,
+ 662.0,
+ 1407.0,
+ 698.0,
+ 295.0,
+ 698.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 695.0,
+ 487.0,
+ 695.0,
+ 487.0,
+ 727.0,
+ 296.0,
+ 727.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 523.0,
+ 695.0,
+ 1404.0,
+ 695.0,
+ 1404.0,
+ 727.0,
+ 523.0,
+ 727.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 723.0,
+ 1405.0,
+ 723.0,
+ 1405.0,
+ 757.0,
+ 293.0,
+ 757.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 750.0,
+ 931.0,
+ 750.0,
+ 931.0,
+ 792.0,
+ 292.0,
+ 792.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 987.0,
+ 750.0,
+ 1407.0,
+ 750.0,
+ 1407.0,
+ 792.0,
+ 987.0,
+ 792.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 785.0,
+ 372.0,
+ 785.0,
+ 372.0,
+ 820.0,
+ 295.0,
+ 820.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 415.0,
+ 785.0,
+ 1038.0,
+ 785.0,
+ 1038.0,
+ 820.0,
+ 415.0,
+ 820.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1115.0,
+ 785.0,
+ 1405.0,
+ 785.0,
+ 1405.0,
+ 820.0,
+ 1115.0,
+ 820.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 814.0,
+ 431.0,
+ 814.0,
+ 431.0,
+ 850.0,
+ 295.0,
+ 850.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 446.0,
+ 814.0,
+ 705.0,
+ 814.0,
+ 705.0,
+ 850.0,
+ 446.0,
+ 850.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 727.0,
+ 814.0,
+ 1046.0,
+ 814.0,
+ 1046.0,
+ 850.0,
+ 727.0,
+ 850.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 322.0,
+ 943.0,
+ 374.0,
+ 943.0,
+ 374.0,
+ 978.0,
+ 322.0,
+ 978.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 401.0,
+ 943.0,
+ 1404.0,
+ 943.0,
+ 1404.0,
+ 978.0,
+ 401.0,
+ 978.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 975.0,
+ 1406.0,
+ 975.0,
+ 1406.0,
+ 1009.0,
+ 294.0,
+ 1009.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1004.0,
+ 1405.0,
+ 1004.0,
+ 1405.0,
+ 1040.0,
+ 294.0,
+ 1040.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1036.0,
+ 1406.0,
+ 1036.0,
+ 1406.0,
+ 1071.0,
+ 295.0,
+ 1071.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1067.0,
+ 1402.0,
+ 1067.0,
+ 1402.0,
+ 1098.0,
+ 296.0,
+ 1098.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1096.0,
+ 775.0,
+ 1096.0,
+ 775.0,
+ 1131.0,
+ 296.0,
+ 1131.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 831.0,
+ 1096.0,
+ 959.0,
+ 1096.0,
+ 959.0,
+ 1131.0,
+ 831.0,
+ 1131.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1052.0,
+ 1096.0,
+ 1402.0,
+ 1096.0,
+ 1402.0,
+ 1131.0,
+ 1052.0,
+ 1131.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1126.0,
+ 499.0,
+ 1126.0,
+ 499.0,
+ 1164.0,
+ 294.0,
+ 1164.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 542.0,
+ 1126.0,
+ 1172.0,
+ 1126.0,
+ 1172.0,
+ 1164.0,
+ 542.0,
+ 1164.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1167.0,
+ 1408.0,
+ 1167.0,
+ 1408.0,
+ 1203.0,
+ 294.0,
+ 1203.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1197.0,
+ 1406.0,
+ 1197.0,
+ 1406.0,
+ 1233.0,
+ 294.0,
+ 1233.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1231.0,
+ 1402.0,
+ 1231.0,
+ 1402.0,
+ 1263.0,
+ 296.0,
+ 1263.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1260.0,
+ 1405.0,
+ 1260.0,
+ 1405.0,
+ 1296.0,
+ 294.0,
+ 1296.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1287.0,
+ 1406.0,
+ 1287.0,
+ 1406.0,
+ 1328.0,
+ 292.0,
+ 1328.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1320.0,
+ 1304.0,
+ 1320.0,
+ 1304.0,
+ 1354.0,
+ 293.0,
+ 1354.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1360.0,
+ 1038.0,
+ 1360.0,
+ 1038.0,
+ 1399.0,
+ 293.0,
+ 1399.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1107.0,
+ 1360.0,
+ 1405.0,
+ 1360.0,
+ 1405.0,
+ 1399.0,
+ 1107.0,
+ 1399.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1391.0,
+ 1405.0,
+ 1391.0,
+ 1405.0,
+ 1428.0,
+ 294.0,
+ 1428.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1423.0,
+ 1405.0,
+ 1423.0,
+ 1405.0,
+ 1459.0,
+ 292.0,
+ 1459.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1454.0,
+ 1405.0,
+ 1454.0,
+ 1405.0,
+ 1491.0,
+ 294.0,
+ 1491.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1487.0,
+ 744.0,
+ 1487.0,
+ 744.0,
+ 1520.0,
+ 296.0,
+ 1520.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1601.0,
+ 1403.0,
+ 1601.0,
+ 1403.0,
+ 1633.0,
+ 296.0,
+ 1633.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1631.0,
+ 1405.0,
+ 1631.0,
+ 1405.0,
+ 1666.0,
+ 295.0,
+ 1666.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1661.0,
+ 1406.0,
+ 1661.0,
+ 1406.0,
+ 1695.0,
+ 292.0,
+ 1695.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1692.0,
+ 1024.0,
+ 1692.0,
+ 1024.0,
+ 1728.0,
+ 295.0,
+ 1728.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1044.0,
+ 1692.0,
+ 1406.0,
+ 1692.0,
+ 1406.0,
+ 1728.0,
+ 1044.0,
+ 1728.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1720.0,
+ 458.0,
+ 1720.0,
+ 458.0,
+ 1763.0,
+ 292.0,
+ 1763.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 514.0,
+ 1720.0,
+ 563.0,
+ 1720.0,
+ 563.0,
+ 1763.0,
+ 514.0,
+ 1763.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 606.0,
+ 1720.0,
+ 1188.0,
+ 1720.0,
+ 1188.0,
+ 1763.0,
+ 606.0,
+ 1763.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1400.0,
+ 1720.0,
+ 1412.0,
+ 1720.0,
+ 1412.0,
+ 1763.0,
+ 1400.0,
+ 1763.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1751.0,
+ 487.0,
+ 1751.0,
+ 487.0,
+ 1788.0,
+ 293.0,
+ 1788.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 507.0,
+ 1751.0,
+ 940.0,
+ 1751.0,
+ 940.0,
+ 1788.0,
+ 507.0,
+ 1788.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1937.0,
+ 545.0,
+ 1937.0,
+ 545.0,
+ 1982.0,
+ 294.0,
+ 1982.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1036.0,
+ 1937.0,
+ 1127.0,
+ 1937.0,
+ 1127.0,
+ 1982.0,
+ 1036.0,
+ 1982.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1168.0,
+ 1937.0,
+ 1408.0,
+ 1937.0,
+ 1408.0,
+ 1982.0,
+ 1168.0,
+ 1982.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1969.0,
+ 723.0,
+ 1969.0,
+ 723.0,
+ 2012.0,
+ 292.0,
+ 2012.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 759.0,
+ 1969.0,
+ 963.0,
+ 1969.0,
+ 963.0,
+ 2012.0,
+ 759.0,
+ 2012.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1102.0,
+ 1969.0,
+ 1152.0,
+ 1969.0,
+ 1152.0,
+ 2012.0,
+ 1102.0,
+ 2012.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1263.0,
+ 1969.0,
+ 1407.0,
+ 1969.0,
+ 1407.0,
+ 2012.0,
+ 1263.0,
+ 2012.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 2001.0,
+ 386.0,
+ 2001.0,
+ 386.0,
+ 2039.0,
+ 296.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 411.0,
+ 2001.0,
+ 1404.0,
+ 2001.0,
+ 1404.0,
+ 2039.0,
+ 411.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 226.0,
+ 1409.0,
+ 226.0,
+ 1409.0,
+ 269.0,
+ 293.0,
+ 269.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 260.0,
+ 1213.0,
+ 260.0,
+ 1213.0,
+ 296.0,
+ 296.0,
+ 296.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 358.0,
+ 1799.0,
+ 740.0,
+ 1799.0,
+ 740.0,
+ 1845.0,
+ 358.0,
+ 1845.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1365.0,
+ 1799.0,
+ 1370.0,
+ 1799.0,
+ 1370.0,
+ 1845.0,
+ 1365.0,
+ 1845.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 357.0,
+ 1843.0,
+ 690.0,
+ 1843.0,
+ 690.0,
+ 1886.0,
+ 357.0,
+ 1886.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 733.0,
+ 1843.0,
+ 797.0,
+ 1843.0,
+ 797.0,
+ 1886.0,
+ 733.0,
+ 1886.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1246.0,
+ 1843.0,
+ 1256.0,
+ 1843.0,
+ 1256.0,
+ 1886.0,
+ 1246.0,
+ 1886.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 357.0,
+ 1886.0,
+ 691.0,
+ 1886.0,
+ 691.0,
+ 1930.0,
+ 357.0,
+ 1930.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 746.0,
+ 1886.0,
+ 814.0,
+ 1886.0,
+ 814.0,
+ 1930.0,
+ 746.0,
+ 1930.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1066.0,
+ 1886.0,
+ 1269.0,
+ 1886.0,
+ 1269.0,
+ 1930.0,
+ 1066.0,
+ 1930.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 3,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1695,
+ 1407,
+ 1695,
+ 1407,
+ 2035,
+ 297,
+ 2035
+ ],
+ "score": 0.983
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 296,
+ 1353,
+ 1405,
+ 1353,
+ 1405,
+ 1478,
+ 296,
+ 1478
+ ],
+ "score": 0.978
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 772,
+ 1046,
+ 772,
+ 1046,
+ 987,
+ 298,
+ 987
+ ],
+ "score": 0.972
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 296,
+ 492,
+ 1405,
+ 492,
+ 1405,
+ 585,
+ 296,
+ 585
+ ],
+ "score": 0.969
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 595,
+ 1046,
+ 595,
+ 1046,
+ 718,
+ 297,
+ 718
+ ],
+ "score": 0.968
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 995,
+ 1050,
+ 995,
+ 1050,
+ 1181,
+ 298,
+ 1181
+ ],
+ "score": 0.966
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1181,
+ 1394,
+ 1181,
+ 1394,
+ 1272,
+ 297,
+ 1272
+ ],
+ "score": 0.963
+ },
+ {
+ "category_id": 3,
+ "poly": [
+ 1074,
+ 631,
+ 1400,
+ 631,
+ 1400,
+ 963,
+ 1074,
+ 963
+ ],
+ "score": 0.961
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1553,
+ 1404,
+ 1553,
+ 1404,
+ 1616,
+ 298,
+ 1616
+ ],
+ "score": 0.958
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 1066,
+ 979,
+ 1404,
+ 979,
+ 1404,
+ 1161,
+ 1066,
+ 1161
+ ],
+ "score": 0.958
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 296,
+ 229,
+ 1405,
+ 229,
+ 1405,
+ 324,
+ 296,
+ 324
+ ],
+ "score": 0.957
+ },
+ {
+ "category_id": 8,
+ "poly": [
+ 672,
+ 1487,
+ 1025,
+ 1487,
+ 1025,
+ 1547,
+ 672,
+ 1547
+ ],
+ "score": 0.951
+ },
+ {
+ "category_id": 8,
+ "poly": [
+ 464,
+ 724,
+ 877,
+ 724,
+ 877,
+ 767,
+ 464,
+ 767
+ ],
+ "score": 0.94
+ },
+ {
+ "category_id": 8,
+ "poly": [
+ 755,
+ 435,
+ 943,
+ 435,
+ 943,
+ 484,
+ 755,
+ 484
+ ],
+ "score": 0.937
+ },
+ {
+ "category_id": 8,
+ "poly": [
+ 463,
+ 334,
+ 1226,
+ 334,
+ 1226,
+ 386,
+ 463,
+ 386
+ ],
+ "score": 0.919
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 299,
+ 74,
+ 813,
+ 74,
+ 813,
+ 105,
+ 299,
+ 105
+ ],
+ "score": 0.911
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 300,
+ 1646,
+ 645,
+ 1646,
+ 645,
+ 1677,
+ 300,
+ 1677
+ ],
+ "score": 0.907
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 299,
+ 1302,
+ 649,
+ 1302,
+ 649,
+ 1333,
+ 299,
+ 1333
+ ],
+ "score": 0.906
+ },
+ {
+ "category_id": 9,
+ "poly": [
+ 1366,
+ 438,
+ 1400,
+ 438,
+ 1400,
+ 468,
+ 1366,
+ 468
+ ],
+ "score": 0.884
+ },
+ {
+ "category_id": 9,
+ "poly": [
+ 1366,
+ 1503,
+ 1400,
+ 1503,
+ 1400,
+ 1533,
+ 1366,
+ 1533
+ ],
+ "score": 0.872
+ },
+ {
+ "category_id": 9,
+ "poly": [
+ 1366,
+ 346,
+ 1400,
+ 346,
+ 1400,
+ 375,
+ 1366,
+ 375
+ ],
+ "score": 0.86
+ },
+ {
+ "category_id": 9,
+ "poly": [
+ 1009,
+ 730,
+ 1042,
+ 730,
+ 1042,
+ 759,
+ 1009,
+ 759
+ ],
+ "score": 0.854
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 858,
+ 2088,
+ 858,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.733
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 329,
+ 393,
+ 1329,
+ 393,
+ 1329,
+ 430,
+ 329,
+ 430
+ ],
+ "score": 0.715
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 671,
+ 1483,
+ 1028,
+ 1483,
+ 1028,
+ 1549,
+ 671,
+ 1549
+ ],
+ "score": 0.94,
+ "latex": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { l e n g t h } } = \\frac { 1 } { 2 } \\left( L - \\sum _ { n = 1 } ^ { N } l _ { n } \\right) ^ { 2 } . } \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 724,
+ 396,
+ 954,
+ 396,
+ 954,
+ 430,
+ 724,
+ 430
+ ],
+ "score": 0.93,
+ "latex": "( T \\leq K _ { p } \\leq 2 T - 1 )"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 462,
+ 723,
+ 877,
+ 723,
+ 877,
+ 768,
+ 462,
+ 768
+ ],
+ "score": 0.92,
+ "latex": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { p r e d } } ^ { \\prime \\prime } = - \\tau \\cdot \\log \\sum _ { p \\in \\mathcal { P } } \\exp \\left( - \\frac { c _ { p } } { \\tau } \\right) , } \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 690,
+ 494,
+ 763,
+ 494,
+ 763,
+ 524,
+ 690,
+ 524
+ ],
+ "score": 0.91,
+ "latex": "p \\in \\mathcal P"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 478,
+ 330,
+ 1220,
+ 330,
+ 1220,
+ 389,
+ 478,
+ 389
+ ],
+ "score": 0.91,
+ "latex": "\\begin{array} { r } { c _ { p } = \\sum _ { k = 1 } ^ { K _ { p } } \\Big ( \\boldsymbol { w } \\cdot \\boldsymbol { \\delta _ { k } } + \\frac { 1 } { F } \\sum _ { f = 1 } ^ { F } | S _ { \\mathrm { g e n } } [ p _ { \\mathrm { g e n } , k } , f ] - S _ { \\mathrm { g t } } [ p _ { \\mathrm { g t } , k } , f ] | \\Big ) . } \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 886,
+ 495,
+ 1095,
+ 495,
+ 1095,
+ 527,
+ 886,
+ 527
+ ],
+ "score": 0.91,
+ "latex": "p _ { \\mathrm { g e n , 1 } } = p _ { \\mathrm { g t , 1 } } = 1"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1144,
+ 494,
+ 1397,
+ 494,
+ 1397,
+ 528,
+ 1144,
+ 528
+ ],
+ "score": 0.9,
+ "latex": "p _ { \\mathrm { g e n } , K _ { p } } = p _ { \\mathrm { g t } , K _ { p } } = T"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 530,
+ 1555,
+ 683,
+ 1555,
+ 683,
+ 1586,
+ 530,
+ 1586
+ ],
+ "score": 0.9,
+ "latex": "\\lambda _ { \\mathrm { l e n g t h } } = 0 . 1"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 370,
+ 773,
+ 475,
+ 773,
+ 475,
+ 801,
+ 370,
+ 801
+ ],
+ "score": 0.9,
+ "latex": "\\tau = 0 . 0 1"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 754,
+ 434,
+ 944,
+ 434,
+ 944,
+ 486,
+ 754,
+ 486
+ ],
+ "score": 0.9,
+ "latex": "\\mathcal { L } _ { \\mathrm { p r e d } } ^ { \\prime } = \\operatorname* { m i n } _ { p \\in \\mathcal { P } } c _ { p } ,"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 378,
+ 836,
+ 463,
+ 836,
+ 463,
+ 862,
+ 378,
+ 862
+ ],
+ "score": 0.89,
+ "latex": "\\tau 0"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 397,
+ 336,
+ 397,
+ 336,
+ 430,
+ 298,
+ 430
+ ],
+ "score": 0.88,
+ "latex": "K _ { p }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 911,
+ 232,
+ 1008,
+ 232,
+ 1008,
+ 261,
+ 911,
+ 261
+ ],
+ "score": 0.88,
+ "latex": "w = 1 . 0"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 803,
+ 444,
+ 803,
+ 444,
+ 836,
+ 298,
+ 836
+ ],
+ "score": 0.88,
+ "latex": "\\lambda _ { \\mathrm { p r e d } } = 1 . 0"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1223,
+ 1555,
+ 1249,
+ 1555,
+ 1249,
+ 1584,
+ 1223,
+ 1584
+ ],
+ "score": 0.88,
+ "latex": "l _ { n }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1127,
+ 263,
+ 1157,
+ 263,
+ 1157,
+ 292,
+ 1127,
+ 292
+ ],
+ "score": 0.87,
+ "latex": "\\delta _ { k }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 1448,
+ 323,
+ 1448,
+ 323,
+ 1477,
+ 298,
+ 1477
+ ],
+ "score": 0.87,
+ "latex": "l _ { n }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1207,
+ 296,
+ 1236,
+ 296,
+ 1236,
+ 326,
+ 1207,
+ 326
+ ],
+ "score": 0.87,
+ "latex": "c _ { p }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 368,
+ 494,
+ 393,
+ 494,
+ 393,
+ 521,
+ 368,
+ 521
+ ],
+ "score": 0.84,
+ "latex": "\\mathcal { P }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 792,
+ 1447,
+ 821,
+ 1447,
+ 821,
+ 1474,
+ 792,
+ 1474
+ ],
+ "score": 0.82,
+ "latex": "N"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 651,
+ 1417,
+ 673,
+ 1417,
+ 673,
+ 1443,
+ 651,
+ 1443
+ ],
+ "score": 0.82,
+ "latex": "L"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 732,
+ 900,
+ 752,
+ 900,
+ 752,
+ 923,
+ 732,
+ 923
+ ],
+ "score": 0.76,
+ "latex": "\\tau"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1315,
+ 1416,
+ 1397,
+ 1416,
+ 1397,
+ 1445,
+ 1315,
+ 1445
+ ],
+ "score": 0.7,
+ "latex": "2 0 0 \\mathrm { H z }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 625,
+ 1451,
+ 643,
+ 1451,
+ 643,
+ 1474,
+ 625,
+ 1474
+ ],
+ "score": 0.66,
+ "latex": "n"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1111.0,
+ 632.0,
+ 1147.0,
+ 632.0,
+ 1147.0,
+ 896.0,
+ 1111.0,
+ 896.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1145.0,
+ 893.0,
+ 1168.0,
+ 893.0,
+ 1168.0,
+ 919.0,
+ 1145.0,
+ 919.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1170.0,
+ 894.0,
+ 1191.0,
+ 894.0,
+ 1191.0,
+ 918.0,
+ 1170.0,
+ 918.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1222.0,
+ 894.0,
+ 1242.0,
+ 894.0,
+ 1242.0,
+ 917.0,
+ 1222.0,
+ 917.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1246.0,
+ 892.0,
+ 1270.0,
+ 892.0,
+ 1270.0,
+ 919.0,
+ 1246.0,
+ 919.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1273.0,
+ 894.0,
+ 1293.0,
+ 894.0,
+ 1293.0,
+ 917.0,
+ 1273.0,
+ 917.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1297.0,
+ 892.0,
+ 1321.0,
+ 892.0,
+ 1321.0,
+ 918.0,
+ 1297.0,
+ 918.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1322.0,
+ 893.0,
+ 1345.0,
+ 893.0,
+ 1345.0,
+ 915.0,
+ 1322.0,
+ 915.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1351.0,
+ 893.0,
+ 1370.0,
+ 893.0,
+ 1370.0,
+ 917.0,
+ 1351.0,
+ 917.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1375.0,
+ 893.0,
+ 1396.0,
+ 893.0,
+ 1396.0,
+ 917.0,
+ 1375.0,
+ 917.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1057.0,
+ 749.5,
+ 1124.0,
+ 749.5,
+ 1124.0,
+ 794.5,
+ 1057.0,
+ 794.5
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1260.25,
+ 931.0,
+ 1275.25,
+ 931.0,
+ 1275.25,
+ 964.0,
+ 1260.25,
+ 964.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1064.0,
+ 976.0,
+ 1407.0,
+ 976.0,
+ 1407.0,
+ 1013.0,
+ 1064.0,
+ 1013.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1063.0,
+ 1006.0,
+ 1404.0,
+ 1006.0,
+ 1404.0,
+ 1044.0,
+ 1063.0,
+ 1044.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1064.0,
+ 1037.0,
+ 1407.0,
+ 1037.0,
+ 1407.0,
+ 1074.0,
+ 1064.0,
+ 1074.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1063.0,
+ 1069.0,
+ 1405.0,
+ 1069.0,
+ 1405.0,
+ 1101.0,
+ 1063.0,
+ 1101.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1064.0,
+ 1102.0,
+ 1405.0,
+ 1102.0,
+ 1405.0,
+ 1132.0,
+ 1064.0,
+ 1132.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1064.0,
+ 1128.0,
+ 1226.0,
+ 1128.0,
+ 1226.0,
+ 1162.0,
+ 1064.0,
+ 1162.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 72.0,
+ 815.0,
+ 72.0,
+ 815.0,
+ 108.0,
+ 295.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1646.0,
+ 644.0,
+ 1646.0,
+ 644.0,
+ 1679.0,
+ 296.0,
+ 1679.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1301.0,
+ 652.0,
+ 1301.0,
+ 652.0,
+ 1337.0,
+ 295.0,
+ 1337.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 839.0,
+ 2085.0,
+ 861.0,
+ 2085.0,
+ 861.0,
+ 2120.0,
+ 839.0,
+ 2120.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1696.0,
+ 1407.0,
+ 1696.0,
+ 1407.0,
+ 1735.0,
+ 294.0,
+ 1735.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1728.0,
+ 1405.0,
+ 1728.0,
+ 1405.0,
+ 1765.0,
+ 294.0,
+ 1765.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1760.0,
+ 1405.0,
+ 1760.0,
+ 1405.0,
+ 1795.0,
+ 295.0,
+ 1795.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1791.0,
+ 1405.0,
+ 1791.0,
+ 1405.0,
+ 1822.0,
+ 294.0,
+ 1822.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1819.0,
+ 1408.0,
+ 1819.0,
+ 1408.0,
+ 1857.0,
+ 291.0,
+ 1857.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1851.0,
+ 1407.0,
+ 1851.0,
+ 1407.0,
+ 1886.0,
+ 294.0,
+ 1886.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1880.0,
+ 1407.0,
+ 1880.0,
+ 1407.0,
+ 1919.0,
+ 294.0,
+ 1919.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1910.0,
+ 1407.0,
+ 1910.0,
+ 1407.0,
+ 1949.0,
+ 294.0,
+ 1949.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1943.0,
+ 1407.0,
+ 1943.0,
+ 1407.0,
+ 1977.0,
+ 294.0,
+ 1977.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1970.0,
+ 1407.0,
+ 1970.0,
+ 1407.0,
+ 2011.0,
+ 292.0,
+ 2011.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 2004.0,
+ 1409.0,
+ 2004.0,
+ 1409.0,
+ 2039.0,
+ 295.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1352.0,
+ 1403.0,
+ 1352.0,
+ 1403.0,
+ 1389.0,
+ 293.0,
+ 1389.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1381.0,
+ 1406.0,
+ 1381.0,
+ 1406.0,
+ 1423.0,
+ 293.0,
+ 1423.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1417.0,
+ 650.0,
+ 1417.0,
+ 650.0,
+ 1449.0,
+ 295.0,
+ 1449.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 674.0,
+ 1417.0,
+ 1314.0,
+ 1417.0,
+ 1314.0,
+ 1449.0,
+ 674.0,
+ 1449.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1398.0,
+ 1417.0,
+ 1407.0,
+ 1417.0,
+ 1407.0,
+ 1449.0,
+ 1398.0,
+ 1449.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1446.0,
+ 297.0,
+ 1446.0,
+ 297.0,
+ 1481.0,
+ 293.0,
+ 1481.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 324.0,
+ 1446.0,
+ 624.0,
+ 1446.0,
+ 624.0,
+ 1481.0,
+ 324.0,
+ 1481.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 644.0,
+ 1446.0,
+ 791.0,
+ 1446.0,
+ 791.0,
+ 1481.0,
+ 644.0,
+ 1481.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 822.0,
+ 1446.0,
+ 1328.0,
+ 1446.0,
+ 1328.0,
+ 1481.0,
+ 822.0,
+ 1481.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 772.0,
+ 369.0,
+ 772.0,
+ 369.0,
+ 806.0,
+ 296.0,
+ 806.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 476.0,
+ 772.0,
+ 1048.0,
+ 772.0,
+ 1048.0,
+ 806.0,
+ 476.0,
+ 806.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 802.0,
+ 297.0,
+ 802.0,
+ 297.0,
+ 836.0,
+ 294.0,
+ 836.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 445.0,
+ 802.0,
+ 1047.0,
+ 802.0,
+ 1047.0,
+ 836.0,
+ 445.0,
+ 836.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 832.0,
+ 377.0,
+ 832.0,
+ 377.0,
+ 868.0,
+ 293.0,
+ 868.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 464.0,
+ 832.0,
+ 1048.0,
+ 832.0,
+ 1048.0,
+ 868.0,
+ 464.0,
+ 868.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 863.0,
+ 1047.0,
+ 863.0,
+ 1047.0,
+ 896.0,
+ 294.0,
+ 896.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 894.0,
+ 731.0,
+ 894.0,
+ 731.0,
+ 929.0,
+ 295.0,
+ 929.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 753.0,
+ 894.0,
+ 1050.0,
+ 894.0,
+ 1050.0,
+ 929.0,
+ 753.0,
+ 929.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 923.0,
+ 1051.0,
+ 923.0,
+ 1051.0,
+ 958.0,
+ 294.0,
+ 958.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 953.0,
+ 1051.0,
+ 953.0,
+ 1051.0,
+ 989.0,
+ 295.0,
+ 989.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 290.0,
+ 487.0,
+ 367.0,
+ 487.0,
+ 367.0,
+ 534.0,
+ 290.0,
+ 534.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 487.0,
+ 689.0,
+ 487.0,
+ 689.0,
+ 534.0,
+ 394.0,
+ 534.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 764.0,
+ 487.0,
+ 885.0,
+ 487.0,
+ 885.0,
+ 534.0,
+ 764.0,
+ 534.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1096.0,
+ 487.0,
+ 1143.0,
+ 487.0,
+ 1143.0,
+ 534.0,
+ 1096.0,
+ 534.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1398.0,
+ 487.0,
+ 1410.0,
+ 487.0,
+ 1410.0,
+ 534.0,
+ 1398.0,
+ 534.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 523.0,
+ 1402.0,
+ 523.0,
+ 1402.0,
+ 557.0,
+ 295.0,
+ 557.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 553.0,
+ 1358.0,
+ 553.0,
+ 1358.0,
+ 589.0,
+ 293.0,
+ 589.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 590.0,
+ 1051.0,
+ 590.0,
+ 1051.0,
+ 632.0,
+ 292.0,
+ 632.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 625.0,
+ 1047.0,
+ 625.0,
+ 1047.0,
+ 660.0,
+ 294.0,
+ 660.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 657.0,
+ 1048.0,
+ 657.0,
+ 1048.0,
+ 687.0,
+ 296.0,
+ 687.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 684.0,
+ 1049.0,
+ 684.0,
+ 1049.0,
+ 718.0,
+ 295.0,
+ 718.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 997.0,
+ 1050.0,
+ 997.0,
+ 1050.0,
+ 1031.0,
+ 295.0,
+ 1031.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1024.0,
+ 1048.0,
+ 1024.0,
+ 1048.0,
+ 1062.0,
+ 294.0,
+ 1062.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1056.0,
+ 1052.0,
+ 1056.0,
+ 1052.0,
+ 1092.0,
+ 292.0,
+ 1092.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1089.0,
+ 1047.0,
+ 1089.0,
+ 1047.0,
+ 1123.0,
+ 295.0,
+ 1123.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1119.0,
+ 1047.0,
+ 1119.0,
+ 1047.0,
+ 1152.0,
+ 293.0,
+ 1152.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1149.0,
+ 1047.0,
+ 1149.0,
+ 1047.0,
+ 1184.0,
+ 294.0,
+ 1184.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1178.0,
+ 1399.0,
+ 1178.0,
+ 1399.0,
+ 1213.0,
+ 292.0,
+ 1213.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1209.0,
+ 1399.0,
+ 1209.0,
+ 1399.0,
+ 1242.0,
+ 296.0,
+ 1242.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1241.0,
+ 676.0,
+ 1241.0,
+ 676.0,
+ 1274.0,
+ 296.0,
+ 1274.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1553.0,
+ 529.0,
+ 1553.0,
+ 529.0,
+ 1588.0,
+ 296.0,
+ 1588.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 684.0,
+ 1553.0,
+ 1222.0,
+ 1553.0,
+ 1222.0,
+ 1588.0,
+ 684.0,
+ 1588.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1250.0,
+ 1553.0,
+ 1404.0,
+ 1553.0,
+ 1404.0,
+ 1588.0,
+ 1250.0,
+ 1588.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1583.0,
+ 984.0,
+ 1583.0,
+ 984.0,
+ 1619.0,
+ 296.0,
+ 1619.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 229.0,
+ 910.0,
+ 229.0,
+ 910.0,
+ 267.0,
+ 293.0,
+ 267.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1009.0,
+ 229.0,
+ 1405.0,
+ 229.0,
+ 1405.0,
+ 267.0,
+ 1009.0,
+ 267.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 262.0,
+ 1126.0,
+ 262.0,
+ 1126.0,
+ 296.0,
+ 295.0,
+ 296.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1158.0,
+ 262.0,
+ 1403.0,
+ 262.0,
+ 1403.0,
+ 296.0,
+ 1158.0,
+ 296.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 289.0,
+ 1206.0,
+ 289.0,
+ 1206.0,
+ 329.0,
+ 291.0,
+ 329.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1237.0,
+ 289.0,
+ 1274.0,
+ 289.0,
+ 1274.0,
+ 329.0,
+ 1237.0,
+ 329.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 337.0,
+ 391.0,
+ 723.0,
+ 391.0,
+ 723.0,
+ 434.0,
+ 337.0,
+ 434.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 955.0,
+ 391.0,
+ 1337.0,
+ 391.0,
+ 1337.0,
+ 434.0,
+ 955.0,
+ 434.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 4,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1212,
+ 1405,
+ 1212,
+ 1405,
+ 1670,
+ 298,
+ 1670
+ ],
+ "score": 0.983
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 611,
+ 1406,
+ 611,
+ 1406,
+ 947,
+ 298,
+ 947
+ ],
+ "score": 0.98
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 295,
+ 1404,
+ 295,
+ 1404,
+ 601,
+ 298,
+ 601
+ ],
+ "score": 0.98
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 957,
+ 1405,
+ 957,
+ 1405,
+ 1202,
+ 298,
+ 1202
+ ],
+ "score": 0.979
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 1680,
+ 1403,
+ 1680,
+ 1403,
+ 1894,
+ 299,
+ 1894
+ ],
+ "score": 0.977
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 302,
+ 1905,
+ 1402,
+ 1905,
+ 1402,
+ 1967,
+ 302,
+ 1967
+ ],
+ "score": 0.935
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 315,
+ 2005,
+ 1385,
+ 2005,
+ 1385,
+ 2034,
+ 315,
+ 2034
+ ],
+ "score": 0.896
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 301,
+ 226,
+ 580,
+ 226,
+ 580,
+ 261,
+ 301,
+ 261
+ ],
+ "score": 0.89
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 300,
+ 75,
+ 813,
+ 75,
+ 813,
+ 104,
+ 300,
+ 104
+ ],
+ "score": 0.875
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 840,
+ 2089,
+ 859,
+ 2089,
+ 859,
+ 2112,
+ 840,
+ 2112
+ ],
+ "score": 0.795
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 330.0,
+ 1998.0,
+ 1386.0,
+ 1998.0,
+ 1386.0,
+ 2040.0,
+ 330.0,
+ 2040.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 222.0,
+ 587.0,
+ 222.0,
+ 587.0,
+ 268.0,
+ 291.0,
+ 268.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 72.0,
+ 815.0,
+ 72.0,
+ 815.0,
+ 108.0,
+ 295.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 840.0,
+ 2087.0,
+ 862.0,
+ 2087.0,
+ 862.0,
+ 2118.0,
+ 840.0,
+ 2118.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1213.0,
+ 1407.0,
+ 1213.0,
+ 1407.0,
+ 1248.0,
+ 295.0,
+ 1248.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1244.0,
+ 1406.0,
+ 1244.0,
+ 1406.0,
+ 1279.0,
+ 295.0,
+ 1279.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1274.0,
+ 1407.0,
+ 1274.0,
+ 1407.0,
+ 1309.0,
+ 295.0,
+ 1309.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1299.0,
+ 1407.0,
+ 1299.0,
+ 1407.0,
+ 1345.0,
+ 291.0,
+ 1345.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1331.0,
+ 1407.0,
+ 1331.0,
+ 1407.0,
+ 1373.0,
+ 291.0,
+ 1373.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1364.0,
+ 1410.0,
+ 1364.0,
+ 1410.0,
+ 1401.0,
+ 292.0,
+ 1401.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1392.0,
+ 1408.0,
+ 1392.0,
+ 1408.0,
+ 1432.0,
+ 292.0,
+ 1432.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1424.0,
+ 1407.0,
+ 1424.0,
+ 1407.0,
+ 1462.0,
+ 293.0,
+ 1462.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1456.0,
+ 1408.0,
+ 1456.0,
+ 1408.0,
+ 1491.0,
+ 295.0,
+ 1491.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1487.0,
+ 1407.0,
+ 1487.0,
+ 1407.0,
+ 1522.0,
+ 295.0,
+ 1522.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1517.0,
+ 1408.0,
+ 1517.0,
+ 1408.0,
+ 1552.0,
+ 295.0,
+ 1552.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1548.0,
+ 1407.0,
+ 1548.0,
+ 1407.0,
+ 1583.0,
+ 295.0,
+ 1583.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1578.0,
+ 1403.0,
+ 1578.0,
+ 1403.0,
+ 1613.0,
+ 295.0,
+ 1613.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1607.0,
+ 1406.0,
+ 1607.0,
+ 1406.0,
+ 1644.0,
+ 293.0,
+ 1644.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1639.0,
+ 1074.0,
+ 1639.0,
+ 1074.0,
+ 1674.0,
+ 295.0,
+ 1674.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 609.0,
+ 1406.0,
+ 609.0,
+ 1406.0,
+ 648.0,
+ 293.0,
+ 648.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 643.0,
+ 1406.0,
+ 643.0,
+ 1406.0,
+ 677.0,
+ 295.0,
+ 677.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 670.0,
+ 1408.0,
+ 670.0,
+ 1408.0,
+ 710.0,
+ 293.0,
+ 710.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 700.0,
+ 1409.0,
+ 700.0,
+ 1409.0,
+ 741.0,
+ 292.0,
+ 741.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 731.0,
+ 1408.0,
+ 731.0,
+ 1408.0,
+ 769.0,
+ 293.0,
+ 769.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 762.0,
+ 1406.0,
+ 762.0,
+ 1406.0,
+ 798.0,
+ 292.0,
+ 798.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 793.0,
+ 1408.0,
+ 793.0,
+ 1408.0,
+ 830.0,
+ 293.0,
+ 830.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 821.0,
+ 1409.0,
+ 821.0,
+ 1409.0,
+ 862.0,
+ 292.0,
+ 862.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 855.0,
+ 1406.0,
+ 855.0,
+ 1406.0,
+ 889.0,
+ 295.0,
+ 889.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 886.0,
+ 1406.0,
+ 886.0,
+ 1406.0,
+ 921.0,
+ 293.0,
+ 921.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 916.0,
+ 1365.0,
+ 916.0,
+ 1365.0,
+ 951.0,
+ 295.0,
+ 951.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 292.0,
+ 1405.0,
+ 292.0,
+ 1405.0,
+ 334.0,
+ 293.0,
+ 334.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 326.0,
+ 1405.0,
+ 326.0,
+ 1405.0,
+ 362.0,
+ 293.0,
+ 362.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 352.0,
+ 1406.0,
+ 352.0,
+ 1406.0,
+ 395.0,
+ 292.0,
+ 395.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 386.0,
+ 1406.0,
+ 386.0,
+ 1406.0,
+ 423.0,
+ 293.0,
+ 423.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 418.0,
+ 1406.0,
+ 418.0,
+ 1406.0,
+ 453.0,
+ 294.0,
+ 453.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 448.0,
+ 1407.0,
+ 448.0,
+ 1407.0,
+ 484.0,
+ 292.0,
+ 484.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 477.0,
+ 1406.0,
+ 477.0,
+ 1406.0,
+ 515.0,
+ 292.0,
+ 515.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 509.0,
+ 1406.0,
+ 509.0,
+ 1406.0,
+ 543.0,
+ 292.0,
+ 543.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 539.0,
+ 1404.0,
+ 539.0,
+ 1404.0,
+ 575.0,
+ 293.0,
+ 575.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 569.0,
+ 1189.0,
+ 569.0,
+ 1189.0,
+ 604.0,
+ 294.0,
+ 604.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 956.0,
+ 1408.0,
+ 956.0,
+ 1408.0,
+ 993.0,
+ 293.0,
+ 993.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 988.0,
+ 1406.0,
+ 988.0,
+ 1406.0,
+ 1022.0,
+ 293.0,
+ 1022.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1018.0,
+ 1406.0,
+ 1018.0,
+ 1406.0,
+ 1052.0,
+ 293.0,
+ 1052.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1050.0,
+ 1406.0,
+ 1050.0,
+ 1406.0,
+ 1084.0,
+ 292.0,
+ 1084.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1077.0,
+ 1405.0,
+ 1077.0,
+ 1405.0,
+ 1115.0,
+ 292.0,
+ 1115.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1108.0,
+ 1409.0,
+ 1108.0,
+ 1409.0,
+ 1148.0,
+ 292.0,
+ 1148.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1141.0,
+ 1405.0,
+ 1141.0,
+ 1405.0,
+ 1174.0,
+ 295.0,
+ 1174.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1172.0,
+ 1099.0,
+ 1172.0,
+ 1099.0,
+ 1203.0,
+ 293.0,
+ 1203.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1678.0,
+ 1404.0,
+ 1678.0,
+ 1404.0,
+ 1715.0,
+ 294.0,
+ 1715.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1710.0,
+ 1405.0,
+ 1710.0,
+ 1405.0,
+ 1747.0,
+ 294.0,
+ 1747.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1740.0,
+ 1404.0,
+ 1740.0,
+ 1404.0,
+ 1778.0,
+ 294.0,
+ 1778.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1774.0,
+ 1405.0,
+ 1774.0,
+ 1405.0,
+ 1807.0,
+ 293.0,
+ 1807.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1803.0,
+ 1405.0,
+ 1803.0,
+ 1405.0,
+ 1837.0,
+ 295.0,
+ 1837.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1831.0,
+ 1407.0,
+ 1831.0,
+ 1407.0,
+ 1869.0,
+ 293.0,
+ 1869.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1864.0,
+ 440.0,
+ 1864.0,
+ 440.0,
+ 1900.0,
+ 294.0,
+ 1900.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1904.0,
+ 1406.0,
+ 1904.0,
+ 1406.0,
+ 1940.0,
+ 296.0,
+ 1940.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1931.0,
+ 1408.0,
+ 1931.0,
+ 1408.0,
+ 1973.0,
+ 295.0,
+ 1973.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 5,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 231,
+ 1404,
+ 231,
+ 1404,
+ 779,
+ 298,
+ 779
+ ],
+ "score": 0.984
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1533,
+ 1404,
+ 1533,
+ 1404,
+ 1810,
+ 298,
+ 1810
+ ],
+ "score": 0.982
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1177,
+ 1405,
+ 1177,
+ 1405,
+ 1452,
+ 298,
+ 1452
+ ],
+ "score": 0.981
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 879,
+ 1404,
+ 879,
+ 1404,
+ 1094,
+ 298,
+ 1094
+ ],
+ "score": 0.98
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1819,
+ 1403,
+ 1819,
+ 1403,
+ 2034,
+ 298,
+ 2034
+ ],
+ "score": 0.977
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 300,
+ 76,
+ 813,
+ 76,
+ 813,
+ 104,
+ 300,
+ 104
+ ],
+ "score": 0.891
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 299,
+ 1126,
+ 677,
+ 1126,
+ 677,
+ 1155,
+ 299,
+ 1155
+ ],
+ "score": 0.886
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 300,
+ 817,
+ 536,
+ 817,
+ 536,
+ 852,
+ 300,
+ 852
+ ],
+ "score": 0.881
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 299,
+ 1483,
+ 474,
+ 1483,
+ 474,
+ 1513,
+ 299,
+ 1513
+ ],
+ "score": 0.875
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 842,
+ 2088,
+ 857,
+ 2088,
+ 857,
+ 2111,
+ 842,
+ 2111
+ ],
+ "score": 0.734
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 805,
+ 232,
+ 840,
+ 232,
+ 840,
+ 261,
+ 805,
+ 261
+ ],
+ "score": 0.88,
+ "latex": "L _ { 1 }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 957,
+ 1973,
+ 1041,
+ 1973,
+ 1041,
+ 2003,
+ 957,
+ 2003
+ ],
+ "score": 0.86,
+ "latex": "\\mathcal { L } _ { \\mathrm { l e n g t h } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 998,
+ 1749,
+ 1053,
+ 1749,
+ 1053,
+ 1778,
+ 998,
+ 1778
+ ],
+ "score": 0.83,
+ "latex": "4 . 4 +"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1131,
+ 1973,
+ 1198,
+ 1973,
+ 1198,
+ 2004,
+ 1131,
+ 2004
+ ],
+ "score": 0.68,
+ "latex": "\\mathcal { L } _ { \\mathrm { p r e d } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 846,
+ 1269,
+ 932,
+ 1269,
+ 932,
+ 1298,
+ 846,
+ 1298
+ ],
+ "score": 0.67,
+ "latex": "2 4 \\mathrm { k H z }"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 72.0,
+ 816.0,
+ 72.0,
+ 816.0,
+ 110.0,
+ 294.0,
+ 110.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1125.0,
+ 681.0,
+ 1125.0,
+ 681.0,
+ 1158.0,
+ 296.0,
+ 1158.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 813.0,
+ 543.0,
+ 813.0,
+ 543.0,
+ 859.0,
+ 291.0,
+ 859.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1478.0,
+ 477.0,
+ 1478.0,
+ 477.0,
+ 1519.0,
+ 293.0,
+ 1519.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 839.0,
+ 2085.0,
+ 861.0,
+ 2085.0,
+ 861.0,
+ 2120.0,
+ 839.0,
+ 2120.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 230.0,
+ 804.0,
+ 230.0,
+ 804.0,
+ 266.0,
+ 296.0,
+ 266.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 841.0,
+ 230.0,
+ 1405.0,
+ 230.0,
+ 1405.0,
+ 266.0,
+ 841.0,
+ 266.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 259.0,
+ 1404.0,
+ 259.0,
+ 1404.0,
+ 299.0,
+ 294.0,
+ 299.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 292.0,
+ 1405.0,
+ 292.0,
+ 1405.0,
+ 327.0,
+ 293.0,
+ 327.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 320.0,
+ 1406.0,
+ 320.0,
+ 1406.0,
+ 359.0,
+ 293.0,
+ 359.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 353.0,
+ 1406.0,
+ 353.0,
+ 1406.0,
+ 386.0,
+ 292.0,
+ 386.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 381.0,
+ 1405.0,
+ 381.0,
+ 1405.0,
+ 416.0,
+ 294.0,
+ 416.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 413.0,
+ 1406.0,
+ 413.0,
+ 1406.0,
+ 448.0,
+ 293.0,
+ 448.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 443.0,
+ 1406.0,
+ 443.0,
+ 1406.0,
+ 478.0,
+ 293.0,
+ 478.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 475.0,
+ 1406.0,
+ 475.0,
+ 1406.0,
+ 510.0,
+ 294.0,
+ 510.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 502.0,
+ 1406.0,
+ 502.0,
+ 1406.0,
+ 540.0,
+ 292.0,
+ 540.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 533.0,
+ 1406.0,
+ 533.0,
+ 1406.0,
+ 571.0,
+ 292.0,
+ 571.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 564.0,
+ 1406.0,
+ 564.0,
+ 1406.0,
+ 602.0,
+ 293.0,
+ 602.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 595.0,
+ 1407.0,
+ 595.0,
+ 1407.0,
+ 630.0,
+ 292.0,
+ 630.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 628.0,
+ 1405.0,
+ 628.0,
+ 1405.0,
+ 659.0,
+ 294.0,
+ 659.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 656.0,
+ 1406.0,
+ 656.0,
+ 1406.0,
+ 691.0,
+ 293.0,
+ 691.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 686.0,
+ 1406.0,
+ 686.0,
+ 1406.0,
+ 725.0,
+ 293.0,
+ 725.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 716.0,
+ 1405.0,
+ 716.0,
+ 1405.0,
+ 756.0,
+ 293.0,
+ 756.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 751.0,
+ 684.0,
+ 751.0,
+ 684.0,
+ 782.0,
+ 296.0,
+ 782.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1533.0,
+ 1406.0,
+ 1533.0,
+ 1406.0,
+ 1568.0,
+ 293.0,
+ 1568.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1563.0,
+ 1406.0,
+ 1563.0,
+ 1406.0,
+ 1602.0,
+ 292.0,
+ 1602.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1594.0,
+ 1406.0,
+ 1594.0,
+ 1406.0,
+ 1630.0,
+ 292.0,
+ 1630.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1626.0,
+ 1404.0,
+ 1626.0,
+ 1404.0,
+ 1661.0,
+ 293.0,
+ 1661.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1656.0,
+ 1405.0,
+ 1656.0,
+ 1405.0,
+ 1692.0,
+ 293.0,
+ 1692.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1686.0,
+ 1406.0,
+ 1686.0,
+ 1406.0,
+ 1723.0,
+ 294.0,
+ 1723.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1715.0,
+ 1408.0,
+ 1715.0,
+ 1408.0,
+ 1753.0,
+ 293.0,
+ 1753.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1746.0,
+ 997.0,
+ 1746.0,
+ 997.0,
+ 1783.0,
+ 295.0,
+ 1783.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1054.0,
+ 1746.0,
+ 1405.0,
+ 1746.0,
+ 1405.0,
+ 1783.0,
+ 1054.0,
+ 1783.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1780.0,
+ 1312.0,
+ 1780.0,
+ 1312.0,
+ 1813.0,
+ 296.0,
+ 1813.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1176.0,
+ 1403.0,
+ 1176.0,
+ 1403.0,
+ 1212.0,
+ 296.0,
+ 1212.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1208.0,
+ 1406.0,
+ 1208.0,
+ 1406.0,
+ 1242.0,
+ 292.0,
+ 1242.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1236.0,
+ 1407.0,
+ 1236.0,
+ 1407.0,
+ 1274.0,
+ 292.0,
+ 1274.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1266.0,
+ 845.0,
+ 1266.0,
+ 845.0,
+ 1304.0,
+ 293.0,
+ 1304.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 933.0,
+ 1266.0,
+ 1405.0,
+ 1266.0,
+ 1405.0,
+ 1304.0,
+ 933.0,
+ 1304.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1298.0,
+ 1405.0,
+ 1298.0,
+ 1405.0,
+ 1334.0,
+ 295.0,
+ 1334.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1325.0,
+ 1406.0,
+ 1325.0,
+ 1406.0,
+ 1368.0,
+ 292.0,
+ 1368.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1360.0,
+ 1405.0,
+ 1360.0,
+ 1405.0,
+ 1397.0,
+ 295.0,
+ 1397.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1389.0,
+ 1407.0,
+ 1389.0,
+ 1407.0,
+ 1425.0,
+ 295.0,
+ 1425.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1420.0,
+ 1390.0,
+ 1420.0,
+ 1390.0,
+ 1456.0,
+ 295.0,
+ 1456.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 877.0,
+ 1408.0,
+ 877.0,
+ 1408.0,
+ 918.0,
+ 292.0,
+ 918.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 911.0,
+ 1405.0,
+ 911.0,
+ 1405.0,
+ 946.0,
+ 294.0,
+ 946.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 942.0,
+ 1402.0,
+ 942.0,
+ 1402.0,
+ 977.0,
+ 294.0,
+ 977.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 973.0,
+ 1404.0,
+ 973.0,
+ 1404.0,
+ 1007.0,
+ 294.0,
+ 1007.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 999.0,
+ 1405.0,
+ 999.0,
+ 1405.0,
+ 1038.0,
+ 291.0,
+ 1038.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1032.0,
+ 1406.0,
+ 1032.0,
+ 1406.0,
+ 1070.0,
+ 293.0,
+ 1070.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1064.0,
+ 1152.0,
+ 1064.0,
+ 1152.0,
+ 1098.0,
+ 293.0,
+ 1098.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1819.0,
+ 1404.0,
+ 1819.0,
+ 1404.0,
+ 1854.0,
+ 294.0,
+ 1854.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1849.0,
+ 1407.0,
+ 1849.0,
+ 1407.0,
+ 1887.0,
+ 293.0,
+ 1887.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1882.0,
+ 1405.0,
+ 1882.0,
+ 1405.0,
+ 1917.0,
+ 294.0,
+ 1917.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1910.0,
+ 1404.0,
+ 1910.0,
+ 1404.0,
+ 1947.0,
+ 292.0,
+ 1947.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1942.0,
+ 1405.0,
+ 1942.0,
+ 1405.0,
+ 1977.0,
+ 294.0,
+ 1977.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1971.0,
+ 956.0,
+ 1971.0,
+ 956.0,
+ 2013.0,
+ 292.0,
+ 2013.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1042.0,
+ 1971.0,
+ 1130.0,
+ 1971.0,
+ 1130.0,
+ 2013.0,
+ 1042.0,
+ 2013.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1199.0,
+ 1971.0,
+ 1408.0,
+ 1971.0,
+ 1408.0,
+ 2013.0,
+ 1199.0,
+ 2013.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 2003.0,
+ 1404.0,
+ 2003.0,
+ 1404.0,
+ 2037.0,
+ 293.0,
+ 2037.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 6,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 5,
+ "poly": [
+ 314,
+ 224,
+ 1378,
+ 224,
+ 1378,
+ 697,
+ 314,
+ 697
+ ],
+ "score": 0.985,
+ "html": " | Stages | 1 Stage | Notes |
| WaveNet (van den Oord et al.,2016) | AAu Ling | × | |
| SampleRNN (Mehri et al., 2017) | AAu | × | not a TTS model |
| Deep Voice (Arik et al.,2017) | Ch APhLing ARAu | × | uses segmentation model |
| WaveRNN (Kalchbrenner etal.,2018) | Ling AR ARAu | × | |
| LPCNet (Valin & Skoglund,2019) | ARAu Cep | × | |
| WaveGlow (Prenger et al.,2019) | MelS FF Au | × | |
| FloWaveNet (Kim et al.,2019) | FAu MelS | × | |
| WaveFlow (Ping et al.,2019b) | ARAu MelS | × | partially autoregressive |
| Par: WaveNet (van den Oord et al.,2018) | Ling FF* →Au | × | distillation |
| ClariNet (Ping et al.,2019a), teacher | ARAu Ch/Ph | √ | |
| ClariNet (Ping etal.,2019a),student | Au Ch/Ph | × | distillation |
| WaveGAN(Donahue et al.,2019) | Au | × | not a TTS model |
| MelGAN (Kumar etal., 2019) | MelsAu | × | |
| Par: WaveGAN(Yamamoto et al.,2020) | Ph AMels FAu | × | |
| AdVoc (Neekhara et al.,2019) | MelsMagS | × | |
| GAN-TTS (Binkowski etal., 2020) | Ling FAu FF | × | |
| Tacotron (Wang et al.,2017) | ChAMels MagS→Au | × | uses Griffin & Lim (1984) |
| Tacotron 2 (Shen et al.,2018) | AMelSAAu Ch | × | |
| Deep Voice 2(Gibiansky et al.,2017) | | × | uses segmentation model |
| DV2 Tacotron (Gibiansky et al.,2017) | ChAMagS AAu | × | |
| Deep Voice 3 (Ping et al.,2018) | Ch AMelS AAu | × | several alternative vocoders |
| TransformerTTS(Li etal.,2019) | Ch→Ph AMelS AAu | × | |
| Flowtron (Valle et al.,2020) | Ch AMelsAu | × | |
| VoiceLoop (Taigman etal.,2017) | Ph ALing → Au | × | |
| GAN Exposure (Guo et al.,2019) | Ph A MelS A Au | × | |
| MelNet (Vasquez & Lewis,2019) | AMelS→Au Ch- | × | |
| ParaNet (Peng et al.,2019) | FF* MelsAu Ch/Ph | × | distillation |
| FastSpeech (Ren et al.,2019) | FAu Ph FF* MelS | × | distillation |
| Flow-TTS (Miao et al.,2020) | Mels Au Ch | | |
| Glow-TTS(Kim et al.,2020) | PhMels Au | × × | |
| Char2wav (Sotelo et al.,2017) | ChALing ARAu | × | end-to-end finetuning |
| EATS (Ours) | Ch/Ph → Au | √ | |
"
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 300,
+ 75,
+ 813,
+ 75,
+ 813,
+ 105,
+ 300,
+ 105
+ ],
+ "score": 0.896
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 835,
+ 2088,
+ 865,
+ 2088,
+ 865,
+ 2112,
+ 835,
+ 2112
+ ],
+ "score": 0.853
+ },
+ {
+ "category_id": 6,
+ "poly": [
+ 296,
+ 1561,
+ 1406,
+ 1561,
+ 1406,
+ 1869,
+ 296,
+ 1869
+ ],
+ "score": 0.694
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1117,
+ 1746,
+ 1184,
+ 1746,
+ 1184,
+ 1777,
+ 1117,
+ 1777
+ ],
+ "score": 0.88,
+ "latex": "( \\mathbf { F } \\mathbf { F } ^ { * } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1048,
+ 1051,
+ 1074,
+ 1051,
+ 1074,
+ 1076,
+ 1048,
+ 1076
+ ],
+ "score": 0.69,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1048,
+ 1017,
+ 1074,
+ 1017,
+ 1074,
+ 1043,
+ 1048,
+ 1043
+ ],
+ "score": 0.68,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1048,
+ 1152,
+ 1074,
+ 1152,
+ 1074,
+ 1178,
+ 1048,
+ 1178
+ ],
+ "score": 0.68,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1048,
+ 1118,
+ 1074,
+ 1118,
+ 1074,
+ 1143,
+ 1048,
+ 1143
+ ],
+ "score": 0.65,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1048,
+ 984,
+ 1073,
+ 984,
+ 1073,
+ 1009,
+ 1048,
+ 1009
+ ],
+ "score": 0.63,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1048,
+ 1186,
+ 1073,
+ 1186,
+ 1073,
+ 1212,
+ 1048,
+ 1212
+ ],
+ "score": 0.63,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1048,
+ 1085,
+ 1073,
+ 1085,
+ 1073,
+ 1110,
+ 1048,
+ 1110
+ ],
+ "score": 0.63,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1049,
+ 883,
+ 1073,
+ 883,
+ 1073,
+ 908,
+ 1049,
+ 908
+ ],
+ "score": 0.62,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1048,
+ 816,
+ 1074,
+ 816,
+ 1074,
+ 842,
+ 1048,
+ 842
+ ],
+ "score": 0.59,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1048,
+ 917,
+ 1073,
+ 917,
+ 1073,
+ 942,
+ 1048,
+ 942
+ ],
+ "score": 0.58,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1049,
+ 715,
+ 1074,
+ 715,
+ 1074,
+ 740,
+ 1049,
+ 740
+ ],
+ "score": 0.58,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1049,
+ 1220,
+ 1073,
+ 1220,
+ 1073,
+ 1245,
+ 1049,
+ 1245
+ ],
+ "score": 0.58,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1048,
+ 849,
+ 1074,
+ 849,
+ 1074,
+ 875,
+ 1048,
+ 875
+ ],
+ "score": 0.57,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1048,
+ 950,
+ 1074,
+ 950,
+ 1074,
+ 975,
+ 1048,
+ 975
+ ],
+ "score": 0.55,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 493,
+ 1624,
+ 544,
+ 1624,
+ 544,
+ 1654,
+ 493,
+ 1654
+ ],
+ "score": 0.49,
+ "latex": "\\mathbf { ( P h ) }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1048,
+ 783,
+ 1074,
+ 783,
+ 1074,
+ 809,
+ 1048,
+ 809
+ ],
+ "score": 0.46,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1049,
+ 1253,
+ 1073,
+ 1253,
+ 1073,
+ 1278,
+ 1049,
+ 1278
+ ],
+ "score": 0.44,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 297,
+ 1624,
+ 353,
+ 1624,
+ 353,
+ 1655,
+ 297,
+ 1655
+ ],
+ "score": 0.43,
+ "latex": "\\mathbf { \\Pi } ( \\mathbf { C h } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 569,
+ 1716,
+ 619,
+ 1716,
+ 619,
+ 1746,
+ 569,
+ 1746
+ ],
+ "score": 0.39,
+ "latex": "\\mathbf { \\Pi } ( \\mathbf { A u } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1049,
+ 682,
+ 1073,
+ 682,
+ 1073,
+ 707,
+ 1049,
+ 707
+ ],
+ "score": 0.29,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1049,
+ 1287,
+ 1073,
+ 1287,
+ 1073,
+ 1312,
+ 1049,
+ 1312
+ ],
+ "score": 0.29,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1049,
+ 1321,
+ 1073,
+ 1321,
+ 1073,
+ 1346,
+ 1049,
+ 1346
+ ],
+ "score": 0.27,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1049,
+ 648,
+ 1073,
+ 648,
+ 1073,
+ 673,
+ 1049,
+ 673
+ ],
+ "score": 0.26,
+ "latex": "\\times"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 73.0,
+ 815.0,
+ 73.0,
+ 815.0,
+ 108.0,
+ 295.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 831.0,
+ 2085.0,
+ 869.0,
+ 2085.0,
+ 869.0,
+ 2124.0,
+ 831.0,
+ 2124.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1563.0,
+ 1403.0,
+ 1563.0,
+ 1403.0,
+ 1595.0,
+ 295.0,
+ 1595.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1593.0,
+ 1406.0,
+ 1593.0,
+ 1406.0,
+ 1628.0,
+ 294.0,
+ 1628.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 354.0,
+ 1623.0,
+ 492.0,
+ 1623.0,
+ 492.0,
+ 1659.0,
+ 354.0,
+ 1659.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 545.0,
+ 1623.0,
+ 1408.0,
+ 1623.0,
+ 1408.0,
+ 1659.0,
+ 545.0,
+ 1659.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1654.0,
+ 1406.0,
+ 1654.0,
+ 1406.0,
+ 1689.0,
+ 294.0,
+ 1689.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1682.0,
+ 1406.0,
+ 1682.0,
+ 1406.0,
+ 1720.0,
+ 293.0,
+ 1720.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1714.0,
+ 568.0,
+ 1714.0,
+ 568.0,
+ 1749.0,
+ 294.0,
+ 1749.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 620.0,
+ 1714.0,
+ 1408.0,
+ 1714.0,
+ 1408.0,
+ 1749.0,
+ 620.0,
+ 1749.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1744.0,
+ 1116.0,
+ 1744.0,
+ 1116.0,
+ 1781.0,
+ 291.0,
+ 1781.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1185.0,
+ 1744.0,
+ 1409.0,
+ 1744.0,
+ 1409.0,
+ 1781.0,
+ 1185.0,
+ 1781.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1775.0,
+ 1405.0,
+ 1775.0,
+ 1405.0,
+ 1811.0,
+ 293.0,
+ 1811.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1807.0,
+ 1406.0,
+ 1807.0,
+ 1406.0,
+ 1843.0,
+ 293.0,
+ 1843.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1836.0,
+ 1248.0,
+ 1836.0,
+ 1248.0,
+ 1872.0,
+ 294.0,
+ 1872.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 21,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 291,
+ 1404,
+ 291,
+ 1404,
+ 504,
+ 297,
+ 504
+ ],
+ "score": 0.979
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 301,
+ 225,
+ 805,
+ 225,
+ 805,
+ 261,
+ 301,
+ 261
+ ],
+ "score": 0.9
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 299,
+ 75,
+ 813,
+ 75,
+ 813,
+ 104,
+ 299,
+ 104
+ ],
+ "score": 0.861
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 835,
+ 2088,
+ 864,
+ 2088,
+ 864,
+ 2112,
+ 835,
+ 2112
+ ],
+ "score": 0.838
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 290.0,
+ 220.0,
+ 812.0,
+ 220.0,
+ 812.0,
+ 270.0,
+ 290.0,
+ 270.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 72.0,
+ 815.0,
+ 72.0,
+ 815.0,
+ 108.0,
+ 295.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 828.0,
+ 2084.0,
+ 871.0,
+ 2084.0,
+ 871.0,
+ 2123.0,
+ 828.0,
+ 2123.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 290.0,
+ 1406.0,
+ 290.0,
+ 1406.0,
+ 328.0,
+ 294.0,
+ 328.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 319.0,
+ 1405.0,
+ 319.0,
+ 1405.0,
+ 356.0,
+ 292.0,
+ 356.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 350.0,
+ 1405.0,
+ 350.0,
+ 1405.0,
+ 387.0,
+ 294.0,
+ 387.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 382.0,
+ 1405.0,
+ 382.0,
+ 1405.0,
+ 417.0,
+ 294.0,
+ 417.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 408.0,
+ 1405.0,
+ 408.0,
+ 1405.0,
+ 450.0,
+ 292.0,
+ 450.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 443.0,
+ 1405.0,
+ 443.0,
+ 1405.0,
+ 479.0,
+ 292.0,
+ 479.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 471.0,
+ 758.0,
+ 471.0,
+ 758.0,
+ 509.0,
+ 294.0,
+ 509.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 22,
+ "width": 1700,
+ "height": 2200
+ }
+ }
+]
\ No newline at end of file
diff --git a/parse/train/ryMxXPFex/images/065a9c2372b0b4a5b85f8dd02bfa2795f995928099bd4ade8e80d09936a39112.jpg b/parse/train/ryMxXPFex/images/065a9c2372b0b4a5b85f8dd02bfa2795f995928099bd4ade8e80d09936a39112.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d96cee84e0709b3a9eb130dffe31f9e4006d57bf
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/065a9c2372b0b4a5b85f8dd02bfa2795f995928099bd4ade8e80d09936a39112.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1baa0b6a8b1576b051d99bf4c0944c0d6c0d14bd0fff7b5f3a5dd9a022f0cf75
+size 9255
diff --git a/parse/train/ryMxXPFex/images/09ccb8ee334093f315e01a9ca8101a59d8617dd5ef61621b586b9750e763de70.jpg b/parse/train/ryMxXPFex/images/09ccb8ee334093f315e01a9ca8101a59d8617dd5ef61621b586b9750e763de70.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..6960968b0e327b393778a1044da97e6f1e0508d2
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/09ccb8ee334093f315e01a9ca8101a59d8617dd5ef61621b586b9750e763de70.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfa82b501254467f270fa174da4c950afcef2180f574cbf34c60c4684974c07c
+size 89971
diff --git a/parse/train/ryMxXPFex/images/0e229301772a577b94a9bbe88faeff7ac8e5f98e0ae2924455834014a8b1ceff.jpg b/parse/train/ryMxXPFex/images/0e229301772a577b94a9bbe88faeff7ac8e5f98e0ae2924455834014a8b1ceff.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..9e577899a3ce4842848df3fdb00b64b24b3d4ee3
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/0e229301772a577b94a9bbe88faeff7ac8e5f98e0ae2924455834014a8b1ceff.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c3e6d785b0de0a2f1a32ea81cb282efe1be0111d1b54972f1ffab8fdfeadcfb
+size 49590
diff --git a/parse/train/ryMxXPFex/images/12bf3aed4575067241ddb111b8c103a998aae39003e7ef2eafe147179b59055b.jpg b/parse/train/ryMxXPFex/images/12bf3aed4575067241ddb111b8c103a998aae39003e7ef2eafe147179b59055b.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..8442b894c3a844489680f4abf72cd8d28f27ea80
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/12bf3aed4575067241ddb111b8c103a998aae39003e7ef2eafe147179b59055b.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1da6caec6f6a6d8d9cbb4456affca93c1f40d2e5459cef2968d313539c6c094
+size 46614
diff --git a/parse/train/ryMxXPFex/images/21185b185df30113ae0dca2151d1059ba699a16cd8f892898e11ddaba0eb1c50.jpg b/parse/train/ryMxXPFex/images/21185b185df30113ae0dca2151d1059ba699a16cd8f892898e11ddaba0eb1c50.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..70e9069dc0dec9933fb5dd26315e695621027024
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/21185b185df30113ae0dca2151d1059ba699a16cd8f892898e11ddaba0eb1c50.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6821e3fe0ce06bc5d468de514dc3344237389738a8dfd211afacbcc835e3246c
+size 3632
diff --git a/parse/train/ryMxXPFex/images/21cae1bb6665cc6952f906354a2905091987df20ef17fcd1caf5da374a35b619.jpg b/parse/train/ryMxXPFex/images/21cae1bb6665cc6952f906354a2905091987df20ef17fcd1caf5da374a35b619.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..fe6e8134d4fc86efb6555be3165e6f7a8a751080
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/21cae1bb6665cc6952f906354a2905091987df20ef17fcd1caf5da374a35b619.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:326af25d9b468693ee1d47ee723921da3ff8b84f6afe0541bdf41161ac4d81ff
+size 6490
diff --git a/parse/train/ryMxXPFex/images/37aaa1d19516286f60a1333a0ab9da262c9b4399ac29c2f3b8b38d5da94f2d12.jpg b/parse/train/ryMxXPFex/images/37aaa1d19516286f60a1333a0ab9da262c9b4399ac29c2f3b8b38d5da94f2d12.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..fde48f26f83e4717464f4447c6a297c9b17412a4
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/37aaa1d19516286f60a1333a0ab9da262c9b4399ac29c2f3b8b38d5da94f2d12.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93b155029e068104a63d02b2b5592646161f6e036ae60d29f0cc97030c80b63f
+size 3291
diff --git a/parse/train/ryMxXPFex/images/4af3036e329efa8685177bc880907491997fe434bf8515f204f0aeaeb2928523.jpg b/parse/train/ryMxXPFex/images/4af3036e329efa8685177bc880907491997fe434bf8515f204f0aeaeb2928523.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..228fe0176698c2edd1e4a98bb29916621428149c
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/4af3036e329efa8685177bc880907491997fe434bf8515f204f0aeaeb2928523.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77441dfd15935d4be1fb13d3e7a57d78449e9bc8568aa96feb37fcbc4512d3e1
+size 2976
diff --git a/parse/train/ryMxXPFex/images/572537d3b027d0d6f1c167773051f2bb8adddd5623dbe2985424ea85efeaafbf.jpg b/parse/train/ryMxXPFex/images/572537d3b027d0d6f1c167773051f2bb8adddd5623dbe2985424ea85efeaafbf.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..a20dd578eafb8ee3b34415686439668d3fd354ae
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/572537d3b027d0d6f1c167773051f2bb8adddd5623dbe2985424ea85efeaafbf.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7583ba75da62e3675d2f5500646517ab7e02db5995b0693aeafe96ffd9142e16
+size 5957
diff --git a/parse/train/ryMxXPFex/images/58712be2d20c439bd065dbd88061861ab1bb51667ae015502cb49b8a20e39b5f.jpg b/parse/train/ryMxXPFex/images/58712be2d20c439bd065dbd88061861ab1bb51667ae015502cb49b8a20e39b5f.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..7d5d0b42362acef52e835d2f4a414ba75682930a
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/58712be2d20c439bd065dbd88061861ab1bb51667ae015502cb49b8a20e39b5f.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc6686dc95a978b854b09cb0b60378cc202249e355255dee26df6dccb29e81fc
+size 14729
diff --git a/parse/train/ryMxXPFex/images/75584855724155dabfedbc22462a0a5098a9b6ff198a556f3c9ad50dc8d02c92.jpg b/parse/train/ryMxXPFex/images/75584855724155dabfedbc22462a0a5098a9b6ff198a556f3c9ad50dc8d02c92.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..ceefa3ed090265f2c3e7015a522888eb8f8f2d82
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/75584855724155dabfedbc22462a0a5098a9b6ff198a556f3c9ad50dc8d02c92.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05a9d92bfd1c1086b117296606eb990eaee8d8ed187de15dba2c05610670eaf2
+size 19834
diff --git a/parse/train/ryMxXPFex/images/80695ecff259b01c4903d90399e9f5448ca13d773121a847a8dfc9d966acbdc2.jpg b/parse/train/ryMxXPFex/images/80695ecff259b01c4903d90399e9f5448ca13d773121a847a8dfc9d966acbdc2.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..04428222834d8ad76e5ebd530cc2eb74da5a747d
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/80695ecff259b01c4903d90399e9f5448ca13d773121a847a8dfc9d966acbdc2.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8fd5e5ea3de5b302dbb4d69fa6633ddb6b75d1b44bb499b7d7e85c23089acfdf
+size 51861
diff --git a/parse/train/ryMxXPFex/images/8c478e93d79a8ef15d994aead218e5bf84c40f0bfcaf4f23047828dd251a28a3.jpg b/parse/train/ryMxXPFex/images/8c478e93d79a8ef15d994aead218e5bf84c40f0bfcaf4f23047828dd251a28a3.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..aa977ba01931b310038695934dcd0e1f2f910c88
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/8c478e93d79a8ef15d994aead218e5bf84c40f0bfcaf4f23047828dd251a28a3.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a1b76c89d0420598e05a3bd495dd19c6e0fa8b924ce9f1b4be8450109d33efc
+size 39123
diff --git a/parse/train/ryMxXPFex/images/8d77533dd665991e71b8e908c0a5fc7bfaa496358fbbb2e156eb1bf1b204ebe1.jpg b/parse/train/ryMxXPFex/images/8d77533dd665991e71b8e908c0a5fc7bfaa496358fbbb2e156eb1bf1b204ebe1.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..65a59e161cd83a93499d3be76be8cfcf5e742009
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/8d77533dd665991e71b8e908c0a5fc7bfaa496358fbbb2e156eb1bf1b204ebe1.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95082181c9f2b44b8d0e95eefd00c641b23f2aebaebf851907529ee57a29adee
+size 9929
diff --git a/parse/train/ryMxXPFex/images/8e5247563d2b09fc1d0bc553ca566977f64d13f4c21d8e15212549b0ae043187.jpg b/parse/train/ryMxXPFex/images/8e5247563d2b09fc1d0bc553ca566977f64d13f4c21d8e15212549b0ae043187.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..1a073fcfd3b2a9c67031608c446253f7b18f6417
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/8e5247563d2b09fc1d0bc553ca566977f64d13f4c21d8e15212549b0ae043187.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c44d0f6b2f4e1f94f90fb7ffaaa18f8b94f9b1007015100a3483e72546737b3
+size 8108
diff --git a/parse/train/ryMxXPFex/images/9b78fe9bba237fc67a0a2c42cf90da2e5689bf0ac2ca99253511f9897b4afcc7.jpg b/parse/train/ryMxXPFex/images/9b78fe9bba237fc67a0a2c42cf90da2e5689bf0ac2ca99253511f9897b4afcc7.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..21fdf04c435eac267275b056ee4ec55a47cffff4
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/9b78fe9bba237fc67a0a2c42cf90da2e5689bf0ac2ca99253511f9897b4afcc7.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cde9c738a4caaf15f744c9805bf214ce0e0f60d3537a4f85991754a0e7c15280
+size 3797
diff --git a/parse/train/ryMxXPFex/images/9c8504931bd54a563f1b26c65ec86fb1bbf51213dedffd3594a5df8511059c6e.jpg b/parse/train/ryMxXPFex/images/9c8504931bd54a563f1b26c65ec86fb1bbf51213dedffd3594a5df8511059c6e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..db2ce0453ef5f27cc9b260382de01056685babce
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/9c8504931bd54a563f1b26c65ec86fb1bbf51213dedffd3594a5df8511059c6e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af312b79ed8e975b1f103153652b02839f69cfe340e07f7037fa583b9e93da1d
+size 9428
diff --git a/parse/train/ryMxXPFex/images/9f97e1df1485360937ab038b487690548123f1fdb8309c7ec723b51f4af12244.jpg b/parse/train/ryMxXPFex/images/9f97e1df1485360937ab038b487690548123f1fdb8309c7ec723b51f4af12244.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..665311370bf598837b912d6e6a7a52ebb953e17a
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/9f97e1df1485360937ab038b487690548123f1fdb8309c7ec723b51f4af12244.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a38c837ac7c02b36f7f6985be2c90ff688b4d3b60c9c1e97bfe22253befef3ef
+size 136152
diff --git a/parse/train/ryMxXPFex/images/a6cdbb93d62a5476f769e59be68d78d415411d0ec43d5a56ab02601aebe89bb2.jpg b/parse/train/ryMxXPFex/images/a6cdbb93d62a5476f769e59be68d78d415411d0ec43d5a56ab02601aebe89bb2.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..08e6469664df5c96729a989eb3b46edcafc588c2
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/a6cdbb93d62a5476f769e59be68d78d415411d0ec43d5a56ab02601aebe89bb2.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82200d633da5bb38e765767bf71072adad4300c10d914ab4565a62d0f9294cf3
+size 19802
diff --git a/parse/train/ryMxXPFex/images/a86041a8a776043fa8dbc51d12b84ac897c8ed3463d73a7ea340fd730fb47b4a.jpg b/parse/train/ryMxXPFex/images/a86041a8a776043fa8dbc51d12b84ac897c8ed3463d73a7ea340fd730fb47b4a.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..1b0904b6f77b33d987b4b051b5a842196842a26c
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/a86041a8a776043fa8dbc51d12b84ac897c8ed3463d73a7ea340fd730fb47b4a.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef6da8299ab89057cbe4a806287c9a3e16760da170c81213fcd0203764bc35fe
+size 15873
diff --git a/parse/train/ryMxXPFex/images/a8c4e4e0ecefba74e0e690a8327442918b903beea5697901329f453b0a6532a5.jpg b/parse/train/ryMxXPFex/images/a8c4e4e0ecefba74e0e690a8327442918b903beea5697901329f453b0a6532a5.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d53fcfdfeb16a34fff0cab1f4bcf627b77288a52
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/a8c4e4e0ecefba74e0e690a8327442918b903beea5697901329f453b0a6532a5.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b00329cc945a61391b280fab598d1549d21d0d966dbe8a4d280914288b21124
+size 6180
diff --git a/parse/train/ryMxXPFex/images/a9a5206e2a3c78e5361074714e55a97c43e95cf0acf074d05cbf85271cd2fe5e.jpg b/parse/train/ryMxXPFex/images/a9a5206e2a3c78e5361074714e55a97c43e95cf0acf074d05cbf85271cd2fe5e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..cba2188adf1b0d6772d5b8aa590fdc7d7ab4d6df
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/a9a5206e2a3c78e5361074714e55a97c43e95cf0acf074d05cbf85271cd2fe5e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4cfef61bf1f458b1dfe035d4d3fc3a08d0bbf702a4cd22c42dddf0100dc77ac1
+size 6350
diff --git a/parse/train/ryMxXPFex/images/ab696401280deb993248aa53b19fd3a26832443a6ad303f1f6252c4510011349.jpg b/parse/train/ryMxXPFex/images/ab696401280deb993248aa53b19fd3a26832443a6ad303f1f6252c4510011349.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..11471a5ef041bbfc1ff8f0196ba3e12e30662b20
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/ab696401280deb993248aa53b19fd3a26832443a6ad303f1f6252c4510011349.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54cca951590969d448d994019d41ac9974f595bf9a1052f955b6726d2bce7c00
+size 19860
diff --git a/parse/train/ryMxXPFex/images/accd552cc3c66a39d90cb2a9064a17971f588004c93bb0575ca634ac92f4b9d0.jpg b/parse/train/ryMxXPFex/images/accd552cc3c66a39d90cb2a9064a17971f588004c93bb0575ca634ac92f4b9d0.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..e891d37af3c60b2ef4e41cf02c670dfe51f7c490
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/accd552cc3c66a39d90cb2a9064a17971f588004c93bb0575ca634ac92f4b9d0.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fc0c314ebb82830135d8db623ebbce2204de5f02555ffd478b5757b37177720
+size 10998
diff --git a/parse/train/ryMxXPFex/images/b48bac40894441b87059d1ac09b03deeb612d566e4a423dc28f0389d6e6fbdbd.jpg b/parse/train/ryMxXPFex/images/b48bac40894441b87059d1ac09b03deeb612d566e4a423dc28f0389d6e6fbdbd.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..2db1a84904fdf4dc6678db0bd14f87a66957a265
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/b48bac40894441b87059d1ac09b03deeb612d566e4a423dc28f0389d6e6fbdbd.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:efc614efc1ba4338a6cab59f0a35c6a02be297dc0dcbb651abd17d1a3d3c46bd
+size 4410
diff --git a/parse/train/ryMxXPFex/images/bb0dcef97d3c50fa37481c93b1417101571a5ec80af74a2f845d569d866fc7e6.jpg b/parse/train/ryMxXPFex/images/bb0dcef97d3c50fa37481c93b1417101571a5ec80af74a2f845d569d866fc7e6.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d99f76ee1906581ddfad7c2b098fc7006b582f4b
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/bb0dcef97d3c50fa37481c93b1417101571a5ec80af74a2f845d569d866fc7e6.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c65e0810d4e23166b77c6032bacbbeb2f8c3c696b833bd7d044290bc2c57f43
+size 5419
diff --git a/parse/train/ryMxXPFex/images/c2c2574b64ea597af4e056f9afa1338b7122ce99cd09219539b7acaaa3e7e531.jpg b/parse/train/ryMxXPFex/images/c2c2574b64ea597af4e056f9afa1338b7122ce99cd09219539b7acaaa3e7e531.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..26b5777ff2e18583a558a5f564820facb7824cfa
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/c2c2574b64ea597af4e056f9afa1338b7122ce99cd09219539b7acaaa3e7e531.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90202e34570ffaa1f92aac7348635639871c72ed54af3e6b282231e5127b5b17
+size 15695
diff --git a/parse/train/ryMxXPFex/images/c9f06001e2ba778f429e974521ee53d12193211e51233a7a31e56f95acef40a7.jpg b/parse/train/ryMxXPFex/images/c9f06001e2ba778f429e974521ee53d12193211e51233a7a31e56f95acef40a7.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..f314e0b23b2771e7b5cac82a429fd1997c0909f4
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/c9f06001e2ba778f429e974521ee53d12193211e51233a7a31e56f95acef40a7.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9df0b19f612f6aa0f87977274328875ba2ad1d70f8643e71960c4a69ad8523e2
+size 7855
diff --git a/parse/train/ryMxXPFex/images/d3e91f2b4a4e767fd4d56f8d42e647a884f23f12f78ba5fcc422bc023914db00.jpg b/parse/train/ryMxXPFex/images/d3e91f2b4a4e767fd4d56f8d42e647a884f23f12f78ba5fcc422bc023914db00.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..0d6b3ccffd458c9a639b59f960f9c970694740a7
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/d3e91f2b4a4e767fd4d56f8d42e647a884f23f12f78ba5fcc422bc023914db00.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:49e11970b4d70ebc1f3cec0e3d97d32c61ca24f1fa35c3af5e84be70dcaaf34b
+size 11418
diff --git a/parse/train/ryMxXPFex/images/d6ac917eb321ea5711957e260982350e99ba8a93588eda3a8b41020b1285e76c.jpg b/parse/train/ryMxXPFex/images/d6ac917eb321ea5711957e260982350e99ba8a93588eda3a8b41020b1285e76c.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..1b1246058b2c49d688493be6af90ccb0d7eab28f
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/d6ac917eb321ea5711957e260982350e99ba8a93588eda3a8b41020b1285e76c.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2a7e60bff52c283c356f6e1631d243a9fd1341cde6d2b3aa88ef4d542138e75
+size 19604
diff --git a/parse/train/ryMxXPFex/images/dbb1052fc7cf2be9c0e20bb4d47e2e4970b5ef1edc64b2d990b7d1ba1d68a2ce.jpg b/parse/train/ryMxXPFex/images/dbb1052fc7cf2be9c0e20bb4d47e2e4970b5ef1edc64b2d990b7d1ba1d68a2ce.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..acbdda0993e051704deb8fa952973ab6e6645b02
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/dbb1052fc7cf2be9c0e20bb4d47e2e4970b5ef1edc64b2d990b7d1ba1d68a2ce.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb3da40085e86ab3959a640d0fa55a3911310e00e487ce3f1bec68ddbbb3ea33
+size 5140
diff --git a/parse/train/ryMxXPFex/images/dc961f3ddf37808526c42e598c8b83e2b321dba9e3fbe36b2e8c1001f42e42cc.jpg b/parse/train/ryMxXPFex/images/dc961f3ddf37808526c42e598c8b83e2b321dba9e3fbe36b2e8c1001f42e42cc.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..1dfbb6a173c22637183b49a4e64bb82f2ae88f06
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/dc961f3ddf37808526c42e598c8b83e2b321dba9e3fbe36b2e8c1001f42e42cc.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:778822bb1f8f435d183eb92980fd0df3b80fdf08e9edb527852f78a8ecd001f0
+size 22214
diff --git a/parse/train/ryMxXPFex/images/ddcc787adef8173e3e8393a077a97d9c0274c46a2e1a5e512fabbaf49ca7e919.jpg b/parse/train/ryMxXPFex/images/ddcc787adef8173e3e8393a077a97d9c0274c46a2e1a5e512fabbaf49ca7e919.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..2335f75b97d7d52367d6ed5d424f2e6ef8764531
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/ddcc787adef8173e3e8393a077a97d9c0274c46a2e1a5e512fabbaf49ca7e919.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57c4a0d95fff01386b215d53ce3c21a1d69253cc9e377d6f30ee7a3bb6227896
+size 29187
diff --git a/parse/train/ryMxXPFex/images/e9b19114f488ccbaf6df0cb2e9f4207ec4ccd25dcdca48796f6c26668f4c80d8.jpg b/parse/train/ryMxXPFex/images/e9b19114f488ccbaf6df0cb2e9f4207ec4ccd25dcdca48796f6c26668f4c80d8.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..20f57dea2209ab3b59c2c25250b6398b0b4df722
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/e9b19114f488ccbaf6df0cb2e9f4207ec4ccd25dcdca48796f6c26668f4c80d8.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d6b5f11788497d6a816bcbd818437ca67852b801acfd7a1ccf8f419ea4faa92
+size 58552
diff --git a/parse/train/ryMxXPFex/images/fae100b47c4e73e345178ad8cba5879cd7c68a3880a9b388e61825bf6e05a0eb.jpg b/parse/train/ryMxXPFex/images/fae100b47c4e73e345178ad8cba5879cd7c68a3880a9b388e61825bf6e05a0eb.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..f1d41e6813acf12434c7bdecd4e2f995c74d2cd8
--- /dev/null
+++ b/parse/train/ryMxXPFex/images/fae100b47c4e73e345178ad8cba5879cd7c68a3880a9b388e61825bf6e05a0eb.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5008e7135a19f701e3a391e6552c6200de53ab6a409f9831b3a03a7dbedf1de1
+size 191562
diff --git a/parse/train/w3x8K0M6sAz/images/00496b03416eee86d41061c56867037edfd20150971023a1fb06382e90a276e9.jpg b/parse/train/w3x8K0M6sAz/images/00496b03416eee86d41061c56867037edfd20150971023a1fb06382e90a276e9.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..15db718e477070fc39c3fadc3dbf8cbff544402d
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/00496b03416eee86d41061c56867037edfd20150971023a1fb06382e90a276e9.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d748c801b05d6675caaf3c887c0b847087fe8d47d44dcd01fcb281ce96bbb093
+size 96975
diff --git a/parse/train/w3x8K0M6sAz/images/13e36310934705c5781a87916fbfdcec3ac3224954790f2a39226e1b68937f5b.jpg b/parse/train/w3x8K0M6sAz/images/13e36310934705c5781a87916fbfdcec3ac3224954790f2a39226e1b68937f5b.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..08163c0043d5570ceab3ab89acc5d03ff93873f4
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/13e36310934705c5781a87916fbfdcec3ac3224954790f2a39226e1b68937f5b.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:606e616e93b2c7cb8ea763046fb4b2164104eef7f21a7ee48c7a8d80b39b0def
+size 27744
diff --git a/parse/train/w3x8K0M6sAz/images/25999ec10387c1b8ccfbc868ce85fd4adb3997f394bbe9fe5415c3fa41a2bab0.jpg b/parse/train/w3x8K0M6sAz/images/25999ec10387c1b8ccfbc868ce85fd4adb3997f394bbe9fe5415c3fa41a2bab0.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..79daecab68980293eafae70aa87e98b45240b5a2
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/25999ec10387c1b8ccfbc868ce85fd4adb3997f394bbe9fe5415c3fa41a2bab0.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5c4b0566610ced036f9239388a9314e4df869a7061b20ebe253299f35b9560b
+size 48222
diff --git a/parse/train/w3x8K0M6sAz/images/3c5d5345b792342769fcf85d480df7dfaf12f9fa1c75f46eec6382748a5fb423.jpg b/parse/train/w3x8K0M6sAz/images/3c5d5345b792342769fcf85d480df7dfaf12f9fa1c75f46eec6382748a5fb423.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..32ddaba755dba511b8059b99586ede837cb38636
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/3c5d5345b792342769fcf85d480df7dfaf12f9fa1c75f46eec6382748a5fb423.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d300c86c784d5e73664034572890d64f1ea4073315e5526ff24bdeaa935f89d4
+size 40214
diff --git a/parse/train/w3x8K0M6sAz/images/40cd7196ef8bd4df78cc3f6c94cf3f60db3b71d839a0cc42b944755b0b8fc868.jpg b/parse/train/w3x8K0M6sAz/images/40cd7196ef8bd4df78cc3f6c94cf3f60db3b71d839a0cc42b944755b0b8fc868.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..6a58786b21dde5d2adca641d2ffb76bfa564b95b
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/40cd7196ef8bd4df78cc3f6c94cf3f60db3b71d839a0cc42b944755b0b8fc868.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78d76283955a723be0945774480967e6ed3a3219eedc8a265bc209bc1c785c01
+size 9432
diff --git a/parse/train/w3x8K0M6sAz/images/42bcb685bc540a9eee6075578f556e2471c48f1ff915ac1c346657af459f487b.jpg b/parse/train/w3x8K0M6sAz/images/42bcb685bc540a9eee6075578f556e2471c48f1ff915ac1c346657af459f487b.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..a46583066b2341dfea96dd66bddd0fc7ef43a62b
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/42bcb685bc540a9eee6075578f556e2471c48f1ff915ac1c346657af459f487b.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b281b3483f2fd3855393647f7ff044a00070d97e4f90ffa6e1130c4ae9c44ef
+size 3351
diff --git a/parse/train/w3x8K0M6sAz/images/550bbdc0f33c4f67fad7538f929d6d585fa604bbd528576f360259da3d083c28.jpg b/parse/train/w3x8K0M6sAz/images/550bbdc0f33c4f67fad7538f929d6d585fa604bbd528576f360259da3d083c28.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..3e16cd7214f25a6613c661122dd20d9f80977266
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/550bbdc0f33c4f67fad7538f929d6d585fa604bbd528576f360259da3d083c28.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b93fd7cc58482b1871e9df9fd0f696c2719baf5c0ed31faa255e68995bbc551f
+size 7222
diff --git a/parse/train/w3x8K0M6sAz/images/6d160bd3b091be39bb9f07d48335845a9ed9ec4d5bd61d729b2de3828a23d52b.jpg b/parse/train/w3x8K0M6sAz/images/6d160bd3b091be39bb9f07d48335845a9ed9ec4d5bd61d729b2de3828a23d52b.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..506601a8b8638f502b295772f9f5e2de42d36459
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/6d160bd3b091be39bb9f07d48335845a9ed9ec4d5bd61d729b2de3828a23d52b.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2889a1ea1bf74ecbdafac3fd5a9c913d3e362befd20c9d693ada5db318cb9924
+size 32138
diff --git a/parse/train/w3x8K0M6sAz/images/97fa692803647ece0e15e8dc9163cbf376d0f9667dfe1963ea76ebb6f2d2a15e.jpg b/parse/train/w3x8K0M6sAz/images/97fa692803647ece0e15e8dc9163cbf376d0f9667dfe1963ea76ebb6f2d2a15e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..72dec8a01335df00fbc8318eb78ef7ce875a4060
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/97fa692803647ece0e15e8dc9163cbf376d0f9667dfe1963ea76ebb6f2d2a15e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:515604367e32bc351de3d82ccb50c9eafa46ee8631b3668528d785a3fe89b332
+size 7945
diff --git a/parse/train/w3x8K0M6sAz/images/c1d448c4035fcc19ece2457264f5788047bdd7c12f8b0da0ea4e63a6a5003357.jpg b/parse/train/w3x8K0M6sAz/images/c1d448c4035fcc19ece2457264f5788047bdd7c12f8b0da0ea4e63a6a5003357.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d1be1027894359e9aa377c029494d4aa74b788da
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/c1d448c4035fcc19ece2457264f5788047bdd7c12f8b0da0ea4e63a6a5003357.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7fb3f729fe6d28db54516cd7904f752cf65d61f658bcd0c9a2fd7ec3144f8d0f
+size 3858
diff --git a/parse/train/w3x8K0M6sAz/images/ced03fa0654fa8e24cf5155f24288804077529fb029dcce36ba9dafdf7019461.jpg b/parse/train/w3x8K0M6sAz/images/ced03fa0654fa8e24cf5155f24288804077529fb029dcce36ba9dafdf7019461.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..84fabd68591013287c6cf9dead1156206e019382
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/ced03fa0654fa8e24cf5155f24288804077529fb029dcce36ba9dafdf7019461.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0b7c051f00c207e7d96df70bc6a12448144c900105749ed76c36ead58a13499
+size 6718
diff --git a/parse/train/w3x8K0M6sAz/images/d1bb9abdc48efa57ccfd11940d646394dc2aae52d1d770844156057bc45f9349.jpg b/parse/train/w3x8K0M6sAz/images/d1bb9abdc48efa57ccfd11940d646394dc2aae52d1d770844156057bc45f9349.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..4d456a12b034e7d11fbe05734a939a585afe79e0
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/d1bb9abdc48efa57ccfd11940d646394dc2aae52d1d770844156057bc45f9349.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac8db5c17bfc43aa40852dca50fb8bacf1fb2acb1499544061327fd709901621
+size 92192
diff --git a/parse/train/w3x8K0M6sAz/images/e827f488f2cb6f4bae1776661058118e29eb4a6a3cf0b0275ff28b885f7fb059.jpg b/parse/train/w3x8K0M6sAz/images/e827f488f2cb6f4bae1776661058118e29eb4a6a3cf0b0275ff28b885f7fb059.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..ca2dc8cdc7f0ff170c2aea5f50bf85486417096d
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/e827f488f2cb6f4bae1776661058118e29eb4a6a3cf0b0275ff28b885f7fb059.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:330f2a01c86696f9eea611e805c9e69be381109e44b4c46410ca2bd7d84a39b8
+size 2980
diff --git a/parse/train/w3x8K0M6sAz/images/f5bc5a21a4d8e008c5065041fd9c8ee0bd160c6729bf1cb6c89a70ee0cd10f49.jpg b/parse/train/w3x8K0M6sAz/images/f5bc5a21a4d8e008c5065041fd9c8ee0bd160c6729bf1cb6c89a70ee0cd10f49.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..cff695dbec835c81635f365dcf1e0682bcb9d773
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/f5bc5a21a4d8e008c5065041fd9c8ee0bd160c6729bf1cb6c89a70ee0cd10f49.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3799dee289568daf4af3c14d3f311ca580fafe2a236e6c8d2b34936756ebc6e
+size 34104
diff --git a/parse/train/w3x8K0M6sAz/images/f912716e23c177a4938148c7a1170b81fdfe86390b15dbfd19bd85e1becda3fb.jpg b/parse/train/w3x8K0M6sAz/images/f912716e23c177a4938148c7a1170b81fdfe86390b15dbfd19bd85e1becda3fb.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..88420d90e7f680962771edd6cd3472c05dcec0ca
--- /dev/null
+++ b/parse/train/w3x8K0M6sAz/images/f912716e23c177a4938148c7a1170b81fdfe86390b15dbfd19bd85e1becda3fb.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29332f6c44cae5a3728bcab31b7024a58dae242ba1a084f59229e1848d183639
+size 31449
diff --git a/parse/train/yhjpeuWepoj/images/06558638152092a6f3c523e045df68d43ae228dd6b46d434df06a6d7d21bf3ce.jpg b/parse/train/yhjpeuWepoj/images/06558638152092a6f3c523e045df68d43ae228dd6b46d434df06a6d7d21bf3ce.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..678069c0289faacfa7017b695f776e3121c431b4
--- /dev/null
+++ b/parse/train/yhjpeuWepoj/images/06558638152092a6f3c523e045df68d43ae228dd6b46d434df06a6d7d21bf3ce.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0bc1e6415313cb1be3df8ca5219f134eec5c35970522852cb7ca1264a392607
+size 3855
diff --git a/parse/train/yhjpeuWepoj/images/0cdd1c9003321317d28d8b4559d03bcd25da016f57b307b23642e9f94c25ffba.jpg b/parse/train/yhjpeuWepoj/images/0cdd1c9003321317d28d8b4559d03bcd25da016f57b307b23642e9f94c25ffba.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..eec2f697044738d37d6f8cd98504bee890949fbe
--- /dev/null
+++ b/parse/train/yhjpeuWepoj/images/0cdd1c9003321317d28d8b4559d03bcd25da016f57b307b23642e9f94c25ffba.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ca94cc8fbb512419bc79465be1dfcf9819a8818ca090ef000aad1ee16f6a933
+size 104417
diff --git a/parse/train/yhjpeuWepoj/images/191bdd0d5b01784262feba9ccaf69858692f4d6b2b63d4147a304c98893c147e.jpg b/parse/train/yhjpeuWepoj/images/191bdd0d5b01784262feba9ccaf69858692f4d6b2b63d4147a304c98893c147e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..5daa0afeeef4bf93b19cc51d37006579eba8d6bb
--- /dev/null
+++ b/parse/train/yhjpeuWepoj/images/191bdd0d5b01784262feba9ccaf69858692f4d6b2b63d4147a304c98893c147e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77d7fbd348064580e773a22e84140c7a16e1ded677ebf78907c804658730bfe6
+size 6153
diff --git a/parse/train/yhjpeuWepoj/images/343b4b429f7a34d25d762b61bf6977f13ba8eaac4be276860b5d2170fa2f18be.jpg b/parse/train/yhjpeuWepoj/images/343b4b429f7a34d25d762b61bf6977f13ba8eaac4be276860b5d2170fa2f18be.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..2861672da9c53582b74fa35a61b0533ba2a89776
--- /dev/null
+++ b/parse/train/yhjpeuWepoj/images/343b4b429f7a34d25d762b61bf6977f13ba8eaac4be276860b5d2170fa2f18be.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06b667f2d972be19444a8ee761ed506d9a7bc8122816d29acfb233492eda0df4
+size 84500
diff --git a/parse/train/yhjpeuWepoj/images/52c51cde4275a75094e8c62c8d4fd77f70977cdffb6dcc8e5fcf1796cfa99e42.jpg b/parse/train/yhjpeuWepoj/images/52c51cde4275a75094e8c62c8d4fd77f70977cdffb6dcc8e5fcf1796cfa99e42.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..fd0349ed5b637b49bd13cd39ab9a93de1f2d6cd1
--- /dev/null
+++ b/parse/train/yhjpeuWepoj/images/52c51cde4275a75094e8c62c8d4fd77f70977cdffb6dcc8e5fcf1796cfa99e42.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ebff2db463bb50cf06f189588773e5ef9e5e1e0e054170f0fabbec6a438acea
+size 31824
diff --git a/parse/train/yhjpeuWepoj/images/573d90daac2f84185a8ae91be8ed3864534da43c9506209ed0e2158432214e91.jpg b/parse/train/yhjpeuWepoj/images/573d90daac2f84185a8ae91be8ed3864534da43c9506209ed0e2158432214e91.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..59de95f80001dc520bcb7af3cc98e5b00646af7d
--- /dev/null
+++ b/parse/train/yhjpeuWepoj/images/573d90daac2f84185a8ae91be8ed3864534da43c9506209ed0e2158432214e91.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ad0b0638f702ca7b60443bffeb7a6a10358d43b704ff55cbb3bd1b7a81baf50
+size 8994
diff --git a/parse/train/yhjpeuWepoj/images/e1748829bba37f8f3b6d43d90dfdbea6be91a7b5cbb5e87bd18f0bf9bd522aa6.jpg b/parse/train/yhjpeuWepoj/images/e1748829bba37f8f3b6d43d90dfdbea6be91a7b5cbb5e87bd18f0bf9bd522aa6.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..e23c03be8cb3364f62ce068198f7d5a381cbc2a0
--- /dev/null
+++ b/parse/train/yhjpeuWepoj/images/e1748829bba37f8f3b6d43d90dfdbea6be91a7b5cbb5e87bd18f0bf9bd522aa6.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3a688e432a7f4563dd54a57c4009798186a8e8886fdf93f08f78318a73fdfb4
+size 32850
diff --git a/parse/train/yhjpeuWepoj/images/fd8c9cc81f0a4968ed7f3124861b6deeb77a38f4952bd4cbfedaa30cfe215f98.jpg b/parse/train/yhjpeuWepoj/images/fd8c9cc81f0a4968ed7f3124861b6deeb77a38f4952bd4cbfedaa30cfe215f98.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..32eb1970faef1acb033dbac1f76e3b6869ab2b35
--- /dev/null
+++ b/parse/train/yhjpeuWepoj/images/fd8c9cc81f0a4968ed7f3124861b6deeb77a38f4952bd4cbfedaa30cfe215f98.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d77193b70c7db0ccdcb4260ade2a7bee13b4b156053fed862ede8b4e73b1adce
+size 46139