diff --git "a/parse/dev/MAMOi89bOL/MAMOi89bOL_middle.json" "b/parse/dev/MAMOi89bOL/MAMOi89bOL_middle.json" new file mode 100644--- /dev/null +++ "b/parse/dev/MAMOi89bOL/MAMOi89bOL_middle.json" @@ -0,0 +1,33218 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 180, + 97, + 431, + 117 + ], + "lines": [ + { + "bbox": [ + 179, + 96, + 432, + 117 + ], + "spans": [ + { + "bbox": [ + 179, + 96, + 432, + 117 + ], + "score": 1.0, + "content": "Masked Autoencoders that Listen", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 130, + 155, + 480, + 179 + ], + "lines": [ + { + "bbox": [ + 174, + 155, + 438, + 168 + ], + "spans": [ + { + "bbox": [ + 174, + 155, + 361, + 168 + ], + "score": 1.0, + "content": "Po-Yao Huang1 Hu Xu1 Juncheng Li2", + "type": "text" + }, + { + "bbox": [ + 367, + 155, + 438, + 168 + ], + "score": 1.0, + "content": "Alexei Baevski1", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 129, + 166, + 480, + 180 + ], + "spans": [ + { + "bbox": [ + 129, + 167, + 361, + 180 + ], + "score": 1.0, + "content": "Michael Auli1 Wojciech Galuba1 Florian Metze1", + "type": "text" + }, + { + "bbox": [ + 367, + 166, + 480, + 180 + ], + "score": 1.0, + "content": "Christoph Feichtenhofer1", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 218, + 189, + 391, + 201 + ], + "lines": [ + { + "bbox": [ + 216, + 186, + 392, + 204 + ], + "spans": [ + { + "bbox": [ + 216, + 186, + 259, + 203 + ], + "score": 1.0, + "content": "1Meta AI", + "type": "text" + }, + { + "bbox": [ + 272, + 187, + 392, + 204 + ], + "score": 1.0, + "content": "2Carnegie Mellon University", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 283, + 220, + 328, + 232 + ], + "lines": [ + { + "bbox": [ + 281, + 219, + 330, + 234 + ], + "spans": [ + { + "bbox": [ + 281, + 219, + 330, + 234 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 236, + 469, + 378 + ], + "lines": [ + { + "bbox": [ + 141, + 236, + 469, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 236, + 469, + 249 + ], + "score": 1.0, + "content": "This paper studies a simple extension of image-based Masked Autoencoders", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 246, + 471, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 471, + 260 + ], + "score": 1.0, + "content": "(MAE) [1] to self-supervised representation learning from audio spectrograms.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 257, + 470, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 470, + 271 + ], + "score": 1.0, + "content": "Following the Transformer encoder-decoder design in MAE, our Audio-MAE", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "score": 1.0, + "content": "first encodes audio spectrogram patches with a high masking ratio, feeding only", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 280, + 470, + 292 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 292 + ], + "score": 1.0, + "content": "the non-masked tokens through encoder layers. The decoder then re-orders and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 290, + 470, + 303 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 470, + 303 + ], + "score": 1.0, + "content": "decodes the encoded context padded with mask tokens, in order to reconstruct", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 302, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 142, + 302, + 470, + 313 + ], + "score": 1.0, + "content": "the input spectrogram. We find it beneficial to incorporate local window atten-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 312, + 470, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 470, + 325 + ], + "score": 1.0, + "content": "tion in the decoder, as audio spectrograms are highly correlated in local time and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 324, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 324, + 469, + 335 + ], + "score": 1.0, + "content": "frequency bands. We then fine-tune the encoder with a lower masking ratio on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 335, + 469, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 335, + 469, + 347 + ], + "score": 1.0, + "content": "target datasets. Empirically, Audio-MAE sets new state-of-the-art performance", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 346, + 469, + 357 + ], + "spans": [ + { + "bbox": [ + 142, + 346, + 469, + 357 + ], + "score": 1.0, + "content": "on six audio and speech classification tasks, outperforming other recent models", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 356, + 469, + 367 + ], + "spans": [ + { + "bbox": [ + 142, + 356, + 469, + 367 + ], + "score": 1.0, + "content": "that use external supervised pre-training. Our code and models is available at", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 367, + 377, + 378 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 377, + 378 + ], + "score": 1.0, + "content": "https://github.com/facebookresearch/AudioMAE.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 389, + 190, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 192, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 192, + 405 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 409, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "score": 1.0, + "content": "Transformers [2] and self-supervised learning [3, 4, 5, 6, 7, 1] are dominating computer vision (CV)", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "and natural language processing (NLP) research. The revolution firstly started in NLP with the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "score": 1.0, + "content": "invention of the Transformer architecture and self-attention [8]. Masked autoencoding with BERT [3]", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 441, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 506, + 455 + ], + "score": 1.0, + "content": "set a new state-of-the-art on various NLP tasks by self-supervised pre-training on large-scale language", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 453, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 465 + ], + "score": 1.0, + "content": "corpus. Similarly in the CV community, Vision Transformers (ViT) [9] have become popular for CV", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 463, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 477 + ], + "score": 1.0, + "content": "tasks, and, for self-supervised image representation learning, Masked Autoencoders (MAE) [1] have", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "brought the CV community closer to the success of BERT in NLP. In addition to the existing masked", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 486, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 480, + 498 + ], + "score": 1.0, + "content": "autoencoders that can read (BERT) or see (MAE), in this work we study those that can listen.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 505, + 622 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "score": 1.0, + "content": "Transformer-based models have recently refreshed leaderboards for audio understanding tasks. For", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "score": 1.0, + "content": "example, AST [10] and MBT [11] improved the audio classification performance on the AudioSet [12],", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 523, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 537 + ], + "score": 1.0, + "content": "Event Sound Classification [13], etc. The key technique behind this is initialization of audio model", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 534, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 548 + ], + "score": 1.0, + "content": "weights with ImageNet pre-trained supervised models (e.g., DeiT [14]) by deflating patch embeddings", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 544, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 559 + ], + "score": 1.0, + "content": "and interpolating positional embeddings for encoding audio spectrograms. However, exploiting", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "ImageNet pre-trained models could be sub-optimal. Unlike initializing video models with weights", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "score": 1.0, + "content": "from image models (e.g., the initial weights of I3D [15] or 3D-ResNets [16] are inflated from", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 577, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 592 + ], + "score": 1.0, + "content": "ImageNet pre-trained image models), there are clear and notable discrepancies between spectrograms", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "representing audio content and natural images. It remains unclear why such heterogeneous image-to-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "audio transfer is useful beyond arguably similar low-level semantics such as shapes of spectrograms", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 611, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 506, + 623 + ], + "score": 1.0, + "content": "and shapes of visual objects. Further, any label bias would inevitably be transferred to audio models.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 505, + 715 + ], + "lines": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "Addressing these concerns, self-supervised audio representation learning has recently attracted much", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "research attention. Based on BEiT [17] that learns to reconstruct image patches or learnt patch", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "tokens, SS-AST [18] extends to the audio domain and exploits spectrograms (akin to 1-channel 2D", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 659, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 674 + ], + "score": 1.0, + "content": "images) and use both contrastive and reconstruction objective as self-supervision. Without using any", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "labels, the key enabler to effective self-supervised representation learning is large-scale pre-training", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 682, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 459, + 695 + ], + "score": 1.0, + "content": "data. In this work we use AudioSet [12] for pre-training, a common dataset containing", + "type": "text" + }, + { + "bbox": [ + 459, + 682, + 472, + 692 + ], + "score": 0.8, + "content": "{ \\sim } 2", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 682, + 506, + 695 + ], + "score": 1.0, + "content": "million", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 692, + 506, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 506, + 706 + ], + "score": 1.0, + "content": "audio recordings. Performing large-scale training with Transformer architectures is challenging as", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 703, + 470, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 470, + 716 + ], + "score": 1.0, + "content": "self-attention in Transformers has quadratic complexity w.r.t. the length of input sequence.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 731, + 385, + 741 + ], + "lines": [ + { + "bbox": [ + 105, + 730, + 387, + 744 + ], + "spans": [ + { + "bbox": [ + 105, + 730, + 387, + 744 + ], + "score": 1.0, + "content": "36th Conference on Neural Information Processing Systems (NeurIPS 2022).", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 180, + 97, + 431, + 117 + ], + "lines": [ + { + "bbox": [ + 179, + 96, + 432, + 117 + ], + "spans": [ + { + "bbox": [ + 179, + 96, + 432, + 117 + ], + "score": 1.0, + "content": "Masked Autoencoders that Listen", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 130, + 155, + 480, + 179 + ], + "lines": [ + { + "bbox": [ + 174, + 155, + 438, + 168 + ], + "spans": [ + { + "bbox": [ + 174, + 155, + 361, + 168 + ], + "score": 1.0, + "content": "Po-Yao Huang1 Hu Xu1 Juncheng Li2", + "type": "text" + }, + { + "bbox": [ + 367, + 155, + 438, + 168 + ], + "score": 1.0, + "content": "Alexei Baevski1", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 129, + 166, + 480, + 180 + ], + "spans": [ + { + "bbox": [ + 129, + 167, + 361, + 180 + ], + "score": 1.0, + "content": "Michael Auli1 Wojciech Galuba1 Florian Metze1", + "type": "text" + }, + { + "bbox": [ + 367, + 166, + 480, + 180 + ], + "score": 1.0, + "content": "Christoph Feichtenhofer1", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 129, + 155, + 480, + 180 + ] + }, + { + "type": "text", + "bbox": [ + 218, + 189, + 391, + 201 + ], + "lines": [ + { + "bbox": [ + 216, + 186, + 392, + 204 + ], + "spans": [ + { + "bbox": [ + 216, + 186, + 259, + 203 + ], + "score": 1.0, + "content": "1Meta AI", + "type": "text" + }, + { + "bbox": [ + 272, + 187, + 392, + 204 + ], + "score": 1.0, + "content": "2Carnegie Mellon University", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3, + "bbox_fs": [ + 216, + 186, + 392, + 204 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 220, + 328, + 232 + ], + "lines": [ + { + "bbox": [ + 281, + 219, + 330, + 234 + ], + "spans": [ + { + "bbox": [ + 281, + 219, + 330, + 234 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 236, + 469, + 378 + ], + "lines": [ + { + "bbox": [ + 141, + 236, + 469, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 236, + 469, + 249 + ], + "score": 1.0, + "content": "This paper studies a simple extension of image-based Masked Autoencoders", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 246, + 471, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 471, + 260 + ], + "score": 1.0, + "content": "(MAE) [1] to self-supervised representation learning from audio spectrograms.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 257, + 470, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 470, + 271 + ], + "score": 1.0, + "content": "Following the Transformer encoder-decoder design in MAE, our Audio-MAE", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "score": 1.0, + "content": "first encodes audio spectrogram patches with a high masking ratio, feeding only", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 280, + 470, + 292 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 292 + ], + "score": 1.0, + "content": "the non-masked tokens through encoder layers. The decoder then re-orders and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 290, + 470, + 303 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 470, + 303 + ], + "score": 1.0, + "content": "decodes the encoded context padded with mask tokens, in order to reconstruct", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 302, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 142, + 302, + 470, + 313 + ], + "score": 1.0, + "content": "the input spectrogram. We find it beneficial to incorporate local window atten-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 312, + 470, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 470, + 325 + ], + "score": 1.0, + "content": "tion in the decoder, as audio spectrograms are highly correlated in local time and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 324, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 324, + 469, + 335 + ], + "score": 1.0, + "content": "frequency bands. We then fine-tune the encoder with a lower masking ratio on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 335, + 469, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 335, + 469, + 347 + ], + "score": 1.0, + "content": "target datasets. Empirically, Audio-MAE sets new state-of-the-art performance", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 346, + 469, + 357 + ], + "spans": [ + { + "bbox": [ + 142, + 346, + 469, + 357 + ], + "score": 1.0, + "content": "on six audio and speech classification tasks, outperforming other recent models", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 356, + 469, + 367 + ], + "spans": [ + { + "bbox": [ + 142, + 356, + 469, + 367 + ], + "score": 1.0, + "content": "that use external supervised pre-training. Our code and models is available at", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 367, + 377, + 378 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 377, + 378 + ], + "score": 1.0, + "content": "https://github.com/facebookresearch/AudioMAE.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11, + "bbox_fs": [ + 141, + 236, + 471, + 378 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 389, + 190, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 192, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 192, + 405 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 409, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "score": 1.0, + "content": "Transformers [2] and self-supervised learning [3, 4, 5, 6, 7, 1] are dominating computer vision (CV)", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "and natural language processing (NLP) research. The revolution firstly started in NLP with the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "score": 1.0, + "content": "invention of the Transformer architecture and self-attention [8]. Masked autoencoding with BERT [3]", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 441, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 506, + 455 + ], + "score": 1.0, + "content": "set a new state-of-the-art on various NLP tasks by self-supervised pre-training on large-scale language", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 453, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 465 + ], + "score": 1.0, + "content": "corpus. Similarly in the CV community, Vision Transformers (ViT) [9] have become popular for CV", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 463, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 477 + ], + "score": 1.0, + "content": "tasks, and, for self-supervised image representation learning, Masked Autoencoders (MAE) [1] have", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "brought the CV community closer to the success of BERT in NLP. In addition to the existing masked", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 486, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 480, + 498 + ], + "score": 1.0, + "content": "autoencoders that can read (BERT) or see (MAE), in this work we study those that can listen.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 408, + 506, + 498 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 505, + 622 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "score": 1.0, + "content": "Transformer-based models have recently refreshed leaderboards for audio understanding tasks. For", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "score": 1.0, + "content": "example, AST [10] and MBT [11] improved the audio classification performance on the AudioSet [12],", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 523, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 537 + ], + "score": 1.0, + "content": "Event Sound Classification [13], etc. The key technique behind this is initialization of audio model", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 534, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 548 + ], + "score": 1.0, + "content": "weights with ImageNet pre-trained supervised models (e.g., DeiT [14]) by deflating patch embeddings", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 544, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 559 + ], + "score": 1.0, + "content": "and interpolating positional embeddings for encoding audio spectrograms. However, exploiting", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "ImageNet pre-trained models could be sub-optimal. Unlike initializing video models with weights", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "score": 1.0, + "content": "from image models (e.g., the initial weights of I3D [15] or 3D-ResNets [16] are inflated from", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 577, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 592 + ], + "score": 1.0, + "content": "ImageNet pre-trained image models), there are clear and notable discrepancies between spectrograms", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "representing audio content and natural images. It remains unclear why such heterogeneous image-to-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "audio transfer is useful beyond arguably similar low-level semantics such as shapes of spectrograms", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 611, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 506, + 623 + ], + "score": 1.0, + "content": "and shapes of visual objects. Further, any label bias would inevitably be transferred to audio models.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 501, + 506, + 623 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 505, + 715 + ], + "lines": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "Addressing these concerns, self-supervised audio representation learning has recently attracted much", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "research attention. Based on BEiT [17] that learns to reconstruct image patches or learnt patch", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "tokens, SS-AST [18] extends to the audio domain and exploits spectrograms (akin to 1-channel 2D", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 659, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 674 + ], + "score": 1.0, + "content": "images) and use both contrastive and reconstruction objective as self-supervision. Without using any", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "labels, the key enabler to effective self-supervised representation learning is large-scale pre-training", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 682, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 459, + 695 + ], + "score": 1.0, + "content": "data. In this work we use AudioSet [12] for pre-training, a common dataset containing", + "type": "text" + }, + { + "bbox": [ + 459, + 682, + 472, + 692 + ], + "score": 0.8, + "content": "{ \\sim } 2", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 682, + 506, + 695 + ], + "score": 1.0, + "content": "million", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 692, + 506, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 506, + 706 + ], + "score": 1.0, + "content": "audio recordings. Performing large-scale training with Transformer architectures is challenging as", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 703, + 470, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 470, + 716 + ], + "score": 1.0, + "content": "self-attention in Transformers has quadratic complexity w.r.t. the length of input sequence.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 627, + 506, + 716 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 0, + 506, + 176 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 105, + 0, + 506, + 176 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 0, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 0, + 506, + 176 + ], + "score": 0.325, + "type": "image", + "image_path": "7b578eaf9796ad13f1638cd79725ff0dce34d674e35e6e70f80938d322ab078e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 105, + 0, + 506, + 58.666666666666664 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 105, + 58.666666666666664, + 506, + 117.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 105, + 117.33333333333333, + 506, + 176.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 181, + 506, + 236 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 181, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 506, + 194 + ], + "score": 1.0, + "content": "Figure 1: Audio-MAE for audio self-supervised learning. An audio recording is first transformed", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 192, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 477, + 204 + ], + "score": 1.0, + "content": "into a spectrogram and split into patches. We embed patches and mask out a large subset", + "type": "text" + }, + { + "bbox": [ + 477, + 192, + 502, + 203 + ], + "score": 0.85, + "content": "( 8 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 192, + 506, + 204 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 268, + 216 + ], + "score": 1.0, + "content": "An encoder then operates on the visible", + "type": "text" + }, + { + "bbox": [ + 269, + 203, + 294, + 214 + ], + "score": 0.88, + "content": "( 2 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 202, + 506, + 216 + ], + "score": 1.0, + "content": "patch embeddings. Finally, a decoder processes the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "order-restored embeddings and mask tokens to reconstruct the input. Audio-MAE is minimizing the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 225, + 495, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 495, + 238 + ], + "score": 1.0, + "content": "mean square error (MSE) on the masked portion of the reconstruction and the input spectrogram.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 505, + 335 + ], + "lines": [ + { + "bbox": [ + 105, + 247, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 259 + ], + "score": 1.0, + "content": "This computational burden has been addressed in different ways. A popular approach is to reduce the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 258, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 270 + ], + "score": 1.0, + "content": "sequence length in self-attention. Various ViT-based architectures have been developed to alleviate", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 269, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 281 + ], + "score": 1.0, + "content": "such issues for image and video understanding. For example, Swin-Transformer [19] only performs", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "score": 1.0, + "content": "local attention within windows that shift across layers. MViT [20] employs pooling attention to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 291, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 505, + 302 + ], + "score": 1.0, + "content": "construct a hierarchy of Transformers where sequence lengths are downsampled. For self-supervised", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 300, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 338, + 315 + ], + "score": 1.0, + "content": "learning, MAE [1] efficiently encodes only a small portion", + "type": "text" + }, + { + "bbox": [ + 338, + 302, + 362, + 312 + ], + "score": 0.85, + "content": "( 2 5 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 300, + 505, + 315 + ], + "score": 1.0, + "content": "of visual patches while the majority", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 312, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 506, + 326 + ], + "score": 1.0, + "content": "of patches is discarded. The simplicity and scalability in MAE make it a promising framework for", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 322, + 252, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 252, + 337 + ], + "score": 1.0, + "content": "large-scale self-supervised learning.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 339, + 505, + 427 + ], + "lines": [ + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "In this work, we study MAE for sound recognition and the unique challenges of the audio domain.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 350, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 506, + 364 + ], + "score": 1.0, + "content": "We present Audio-MAE (Fig. 1) as unified and scalable framework for learning self-supervised audio", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "score": 1.0, + "content": "representations. Similar to MAE, it is composed of a pair of a Transformer encoder and decoder.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "score": 1.0, + "content": "Sound is first transformed and embedded into spectrogram patches. Before feeding them into the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "Transformer encoder, we mask and discard the majority and only feed a small number of non-masked", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 505, + 406 + ], + "score": 1.0, + "content": "embeddings into the encoder for efficient encoding. After padding encoded patches with learnable", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 406, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 417 + ], + "score": 1.0, + "content": "embeddings to represent masked patches, it then restores the order of these patches in frequency and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 416, + 488, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 488, + 428 + ], + "score": 1.0, + "content": "time and propagates them through a Transformer decoder to reconstruct the audio spectrogram.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 433, + 505, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 445 + ], + "score": 1.0, + "content": "Different from image patches, spectrogram patches are comparably local-correlated. For example,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 442, + 507, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 507, + 456 + ], + "score": 1.0, + "content": "formants, the vocal tract resonances, are typically grouped and continuous locally in the spectrogram.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "score": 1.0, + "content": "The location in frequency and time embeds essential information that determines the semantics of a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "spectrogram patch and how it sounds like. To this end, we further investigate using localized attention", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "score": 1.0, + "content": "and a hybrid architecture in the Transformer decoder to properly decode for reconstruction. This", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 487, + 419, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 419, + 499 + ], + "score": 1.0, + "content": "simple-yet-effective upgrade leads to improved performance for Audio-MAE.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 505, + 517 + ], + "score": 1.0, + "content": "Similar to MAE for images, we minimize the patch-normalized mean square error. At the fine-tuning", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "stage, we discard the decoder and fine-tune the encoder with patch-masking. Empirically, Audio-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "MAE sets a new state-of-the-art performance on six audio and speech classification tasks. It is the first", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 534, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 550 + ], + "score": 1.0, + "content": "audio-only self-supervised model that achieves state-of-the-art mAP on AudioSet-2M, outperforming", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 547, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 506, + 559 + ], + "score": 1.0, + "content": "other recent models with external supervision. We further provide the visualization and audible", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 558, + 443, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 443, + 570 + ], + "score": 1.0, + "content": "examples to qualitatively demonstrate the effectiveness of the Audio-MAE decoder.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5 + }, + { + "type": "title", + "bbox": [ + 107, + 584, + 197, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 198, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 198, + 599 + ], + "score": 1.0, + "content": "2 Related Work", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "Visual masked pre-training. Masked/Denoising autoencoders [21, 22, 3] are a general represen-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 506, + 636 + ], + "score": 1.0, + "content": "tation learning methodology by reconstructing source from masked or corrupted inputs. In CV,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 505, + 647 + ], + "score": 1.0, + "content": "visual masked pre-training has made recent progress [23, 24, 1, 20]. Based on ViT [9] that applies", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "Transformers to image patches, BEiT [17] and MAE [1] present masked image modeling frameworks.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 655, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 669 + ], + "score": 1.0, + "content": "BEiT [17] learns to predict discrete visual tokens generated by VAE [25] in masked patches. MAE [1]", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "reduces sequence length by masking a large portion of image patches randomly and encoding only", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "non-masked ones for reconstruction of pixel color information. MaskFeat [20] studies features for", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "masked pre-training and finds that Histograms of Oriented Gradients (HoG) [26], which are in turn", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "related to spectrogram features, perform strongly for image and video classification models. Our", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 710, + 458, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 458, + 723 + ], + "score": 1.0, + "content": "work extends the MAE framework for representation learning with audio spectrograms.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 41.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 0, + 506, + 176 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 105, + 0, + 506, + 176 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 0, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 0, + 506, + 176 + ], + "score": 0.325, + "type": "image", + "image_path": "7b578eaf9796ad13f1638cd79725ff0dce34d674e35e6e70f80938d322ab078e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 105, + 0, + 506, + 58.666666666666664 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 105, + 58.666666666666664, + 506, + 117.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 105, + 117.33333333333333, + 506, + 176.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 181, + 506, + 236 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 181, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 506, + 194 + ], + "score": 1.0, + "content": "Figure 1: Audio-MAE for audio self-supervised learning. An audio recording is first transformed", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 192, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 477, + 204 + ], + "score": 1.0, + "content": "into a spectrogram and split into patches. We embed patches and mask out a large subset", + "type": "text" + }, + { + "bbox": [ + 477, + 192, + 502, + 203 + ], + "score": 0.85, + "content": "( 8 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 192, + 506, + 204 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 268, + 216 + ], + "score": 1.0, + "content": "An encoder then operates on the visible", + "type": "text" + }, + { + "bbox": [ + 269, + 203, + 294, + 214 + ], + "score": 0.88, + "content": "( 2 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 202, + 506, + 216 + ], + "score": 1.0, + "content": "patch embeddings. Finally, a decoder processes the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "order-restored embeddings and mask tokens to reconstruct the input. Audio-MAE is minimizing the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 225, + 495, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 495, + 238 + ], + "score": 1.0, + "content": "mean square error (MSE) on the masked portion of the reconstruction and the input spectrogram.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 505, + 335 + ], + "lines": [ + { + "bbox": [ + 105, + 247, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 259 + ], + "score": 1.0, + "content": "This computational burden has been addressed in different ways. A popular approach is to reduce the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 258, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 270 + ], + "score": 1.0, + "content": "sequence length in self-attention. Various ViT-based architectures have been developed to alleviate", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 269, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 281 + ], + "score": 1.0, + "content": "such issues for image and video understanding. For example, Swin-Transformer [19] only performs", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "score": 1.0, + "content": "local attention within windows that shift across layers. MViT [20] employs pooling attention to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 291, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 505, + 302 + ], + "score": 1.0, + "content": "construct a hierarchy of Transformers where sequence lengths are downsampled. For self-supervised", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 300, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 338, + 315 + ], + "score": 1.0, + "content": "learning, MAE [1] efficiently encodes only a small portion", + "type": "text" + }, + { + "bbox": [ + 338, + 302, + 362, + 312 + ], + "score": 0.85, + "content": "( 2 5 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 300, + 505, + 315 + ], + "score": 1.0, + "content": "of visual patches while the majority", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 312, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 506, + 326 + ], + "score": 1.0, + "content": "of patches is discarded. The simplicity and scalability in MAE make it a promising framework for", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 322, + 252, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 252, + 337 + ], + "score": 1.0, + "content": "large-scale self-supervised learning.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 247, + 506, + 337 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 339, + 505, + 427 + ], + "lines": [ + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "In this work, we study MAE for sound recognition and the unique challenges of the audio domain.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 350, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 506, + 364 + ], + "score": 1.0, + "content": "We present Audio-MAE (Fig. 1) as unified and scalable framework for learning self-supervised audio", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "score": 1.0, + "content": "representations. Similar to MAE, it is composed of a pair of a Transformer encoder and decoder.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "score": 1.0, + "content": "Sound is first transformed and embedded into spectrogram patches. Before feeding them into the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "Transformer encoder, we mask and discard the majority and only feed a small number of non-masked", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 505, + 406 + ], + "score": 1.0, + "content": "embeddings into the encoder for efficient encoding. After padding encoded patches with learnable", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 406, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 417 + ], + "score": 1.0, + "content": "embeddings to represent masked patches, it then restores the order of these patches in frequency and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 416, + 488, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 488, + 428 + ], + "score": 1.0, + "content": "time and propagates them through a Transformer decoder to reconstruct the audio spectrogram.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 339, + 506, + 428 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 433, + 505, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 445 + ], + "score": 1.0, + "content": "Different from image patches, spectrogram patches are comparably local-correlated. For example,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 442, + 507, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 507, + 456 + ], + "score": 1.0, + "content": "formants, the vocal tract resonances, are typically grouped and continuous locally in the spectrogram.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "score": 1.0, + "content": "The location in frequency and time embeds essential information that determines the semantics of a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "spectrogram patch and how it sounds like. To this end, we further investigate using localized attention", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "score": 1.0, + "content": "and a hybrid architecture in the Transformer decoder to properly decode for reconstruction. This", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 487, + 419, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 419, + 499 + ], + "score": 1.0, + "content": "simple-yet-effective upgrade leads to improved performance for Audio-MAE.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 432, + 507, + 499 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 505, + 517 + ], + "score": 1.0, + "content": "Similar to MAE for images, we minimize the patch-normalized mean square error. At the fine-tuning", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "stage, we discard the decoder and fine-tune the encoder with patch-masking. Empirically, Audio-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "MAE sets a new state-of-the-art performance on six audio and speech classification tasks. It is the first", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 534, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 550 + ], + "score": 1.0, + "content": "audio-only self-supervised model that achieves state-of-the-art mAP on AudioSet-2M, outperforming", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 547, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 506, + 559 + ], + "score": 1.0, + "content": "other recent models with external supervision. We further provide the visualization and audible", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 558, + 443, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 443, + 570 + ], + "score": 1.0, + "content": "examples to qualitatively demonstrate the effectiveness of the Audio-MAE decoder.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 501, + 506, + 570 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 584, + 197, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 198, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 198, + 599 + ], + "score": 1.0, + "content": "2 Related Work", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "Visual masked pre-training. Masked/Denoising autoencoders [21, 22, 3] are a general represen-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 506, + 636 + ], + "score": 1.0, + "content": "tation learning methodology by reconstructing source from masked or corrupted inputs. In CV,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 505, + 647 + ], + "score": 1.0, + "content": "visual masked pre-training has made recent progress [23, 24, 1, 20]. Based on ViT [9] that applies", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "Transformers to image patches, BEiT [17] and MAE [1] present masked image modeling frameworks.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 655, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 669 + ], + "score": 1.0, + "content": "BEiT [17] learns to predict discrete visual tokens generated by VAE [25] in masked patches. MAE [1]", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "reduces sequence length by masking a large portion of image patches randomly and encoding only", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "non-masked ones for reconstruction of pixel color information. MaskFeat [20] studies features for", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "masked pre-training and finds that Histograms of Oriented Gradients (HoG) [26], which are in turn", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "related to spectrogram features, perform strongly for image and video classification models. Our", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 710, + 458, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 458, + 723 + ], + "score": 1.0, + "content": "work extends the MAE framework for representation learning with audio spectrograms.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 612, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "Out-of-domain pre-training for audio. Transferring ImageNet supervised pre-trained ViT [9]", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 84, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 506, + 96 + ], + "score": 1.0, + "content": "or ResNet [27] has become a popular practice for audio models [10, 28, 11, 29, 30, 31]. After", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "pre-training, these models operate over audio spectrograms by deflating from 3-channels (RGB)", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 506, + 118 + ], + "score": 1.0, + "content": "into 1-channel (spectrogram) in the pre-trained patch embedding in ViT and employing the rest of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 506, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 506, + 130 + ], + "score": 1.0, + "content": "the transformer blocks on top. For example, HTS-AT [29] encodes spectrograms with hierarchical", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "Transformer initialized from the Swin Transformer [19]. MBT [11] uses ImageNet-21K pre-trained", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 136, + 506, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 152 + ], + "score": 1.0, + "content": "ViT; AST [10] and PaSST [28] employ DeiT [14] as the Transformer backbone. Without using", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "out-of-domain (non-audio) data, the proposed Audio-MAE focuses on audio-only self-supervised", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 161, + 211, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 211, + 172 + ], + "score": 1.0, + "content": "pre-training from scratch.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 179, + 506, + 322 + ], + "lines": [ + { + "bbox": [ + 106, + 179, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 179, + 505, + 192 + ], + "score": 1.0, + "content": "In-domain pre-training for audio. Existing in-domain (i.e., audio-only) self-supervised methods", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 190, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 506, + 204 + ], + "score": 1.0, + "content": "can be broadly categorized by the input signal type (e.g., raw waveform [32, 33, 34], frame-level", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 200, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 104, + 200, + 506, + 215 + ], + "score": 1.0, + "content": "features [35, 36, 37], or spectrogram patches [18, 38]); and the objective used for self-supervision", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 212, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 104, + 212, + 506, + 226 + ], + "score": 1.0, + "content": "(e.g., contrastive [39, 33, 40, 41, 35] or prediction/reconstruction [18, 34, 37, 36]). For example,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 223, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 237 + ], + "score": 1.0, + "content": "wav2vec 2.0 [33] takes raw waveform as inputs and exploits contrastive learning to discriminate", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 234, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 246 + ], + "score": 1.0, + "content": "contextualized representations in different time segments. Mockingjay [42] proposed a masked", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 245, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 506, + 257 + ], + "score": 1.0, + "content": "acoustic model pretext task to reconstruct frame-level Mel-features of masked time frames. SS-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 257, + 506, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 506, + 268 + ], + "score": 1.0, + "content": "AST [18] is the closest work to Audio-MAE and is our main benchmark. Inspired by the success of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "BERT [3], SS-AST proposed a self-supervised learning method which operates over spectrogram", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 278, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 506, + 291 + ], + "score": 1.0, + "content": "patches and employs joint contrastive and reconstructive objectives on masked patches. These", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "score": 1.0, + "content": "previous methods generate audio representations by encoding full-view of both masked and non-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 299, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 506, + 313 + ], + "score": 1.0, + "content": "masked time or spectrogram segments for self-supervised pre-training. In contrast, Audio-MAE", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 312, + 313, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 313, + 323 + ], + "score": 1.0, + "content": "encodes only the non-masked spectrogram patches.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 327, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "score": 1.0, + "content": "Our work is done independently and concurrently with [38, 43, 44] related methods. We also compare", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 338, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 506, + 350 + ], + "score": 1.0, + "content": "our model to these concurrent works in the experiments and showcase the superiority of Audio-MAE.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "title", + "bbox": [ + 107, + 366, + 346, + 380 + ], + "lines": [ + { + "bbox": [ + 104, + 365, + 347, + 383 + ], + "spans": [ + { + "bbox": [ + 104, + 365, + 347, + 383 + ], + "score": 1.0, + "content": "3 Audio Masked Autoencoders (Audio-MAE)", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 504, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "Audio-MAE is a conceptually simple extension of MAE to learn self-supervised representations from", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 403, + 484, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 484, + 416 + ], + "score": 1.0, + "content": "audio spectrograms. Fig. 1 depicts an overview. The details of each component are as follows.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 419, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 419, + 507, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 507, + 432 + ], + "score": 1.0, + "content": "Spectrogram Patch Embeddings. Following [10, 18], we transform audio recordings into Mel-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "score": 1.0, + "content": "spectrograms and divide them into non-overlapped regular grid patches. These patches are then", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 441, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 454 + ], + "score": 1.0, + "content": "flattened and embedded by a linear projection. Similar to MAE [1], we add fixed sinusoidal positional", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 452, + 260, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 260, + 464 + ], + "score": 1.0, + "content": "embeddings to the embedded patches.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "image", + "bbox": [ + 112, + 476, + 498, + 543 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 476, + 498, + 543 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 476, + 498, + 543 + ], + "spans": [ + { + "bbox": [ + 112, + 476, + 498, + 543 + ], + "score": 0.969, + "type": "image", + "image_path": "31a937966647bc9c70a9fc197b7e10a2aa38720c7ea2a23229a1c5f045f9bb94.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 112, + 476, + 498, + 498.3333333333333 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 112, + 498.3333333333333, + 498, + 520.6666666666666 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 112, + 520.6666666666666, + 498, + 543.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 172, + 549, + 437, + 561 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 172, + 547, + 439, + 564 + ], + "spans": [ + { + "bbox": [ + 172, + 547, + 439, + 564 + ], + "score": 1.0, + "content": "Figure 2: Audio-MAE’s masking strategies on Mel-spectrograms.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + } + ], + "index": 33.0 + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 505, + 641 + ], + "lines": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "Masking Strategies. Audio-MAE masks out a large subset of spectrogram patches. As a spectrogram", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 585, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 598 + ], + "score": 1.0, + "content": "can be viewed as a 2D representation of time and frequency components of a sound, it is reasonable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "to explore treating time and frequency differently during masking. In this work, we investigate both", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "score": 1.0, + "content": "the unstructured (i.e., random masking without any prior) and structured (i.e., randomly masking a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 617, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 104, + 617, + 245, + 632 + ], + "score": 1.0, + "content": "portion of time, frequency, or time", + "type": "text" + }, + { + "bbox": [ + 245, + 619, + 253, + 628 + ], + "score": 0.33, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 617, + 506, + 632 + ], + "score": 1.0, + "content": "frequency of a spectrogram) in the pre-training and fine-tuning", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 628, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 480, + 642 + ], + "score": 1.0, + "content": "phase. Illustrative examples are shown in Fig. 2. We show masked regions with dark overlay.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "The masking mechanism, as introduced in MAE [1], is the key ingredient for efficient self-supervised", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 656, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 506, + 668 + ], + "score": 1.0, + "content": "learning. For a input patch sequence, this can be regarded as a Bernoulli process where each patch is", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 237, + 680 + ], + "score": 1.0, + "content": "masked/dropped with probability", + "type": "text" + }, + { + "bbox": [ + 238, + 669, + 244, + 679 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "(masking ratio). Masking reduces input patch sequence length and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 679, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 505, + 690 + ], + "score": 1.0, + "content": "encourages learning global, contextualized representations from limited “visible” patches. We observe", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 276, + 701 + ], + "score": 1.0, + "content": "that akin to images, a large masking rate (", + "type": "text" + }, + { + "bbox": [ + 276, + 689, + 296, + 699 + ], + "score": 0.86, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "in our experiments for spectrogram patches, which", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 156, + 713 + ], + "score": 1.0, + "content": "is similar to", + "type": "text" + }, + { + "bbox": [ + 156, + 700, + 176, + 711 + ], + "score": 0.88, + "content": "7 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "in MAE for images) is feasible for learning self-supervised audio representations.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 709, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 222, + 724 + ], + "score": 1.0, + "content": "Unlike BERT [3] that uses", + "type": "text" + }, + { + "bbox": [ + 222, + 711, + 242, + 721 + ], + "score": 0.85, + "content": "15 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 709, + 505, + 724 + ], + "score": 1.0, + "content": "masking rate for self-supervised learning in NLP, most of the", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "Out-of-domain pre-training for audio. Transferring ImageNet supervised pre-trained ViT [9]", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 84, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 506, + 96 + ], + "score": 1.0, + "content": "or ResNet [27] has become a popular practice for audio models [10, 28, 11, 29, 30, 31]. After", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "pre-training, these models operate over audio spectrograms by deflating from 3-channels (RGB)", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 506, + 118 + ], + "score": 1.0, + "content": "into 1-channel (spectrogram) in the pre-trained patch embedding in ViT and employing the rest of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 506, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 506, + 130 + ], + "score": 1.0, + "content": "the transformer blocks on top. For example, HTS-AT [29] encodes spectrograms with hierarchical", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "Transformer initialized from the Swin Transformer [19]. MBT [11] uses ImageNet-21K pre-trained", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 136, + 506, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 152 + ], + "score": 1.0, + "content": "ViT; AST [10] and PaSST [28] employ DeiT [14] as the Transformer backbone. Without using", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "out-of-domain (non-audio) data, the proposed Audio-MAE focuses on audio-only self-supervised", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 161, + 211, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 211, + 172 + ], + "score": 1.0, + "content": "pre-training from scratch.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 73, + 506, + 172 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 179, + 506, + 322 + ], + "lines": [ + { + "bbox": [ + 106, + 179, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 179, + 505, + 192 + ], + "score": 1.0, + "content": "In-domain pre-training for audio. Existing in-domain (i.e., audio-only) self-supervised methods", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 190, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 506, + 204 + ], + "score": 1.0, + "content": "can be broadly categorized by the input signal type (e.g., raw waveform [32, 33, 34], frame-level", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 200, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 104, + 200, + 506, + 215 + ], + "score": 1.0, + "content": "features [35, 36, 37], or spectrogram patches [18, 38]); and the objective used for self-supervision", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 212, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 104, + 212, + 506, + 226 + ], + "score": 1.0, + "content": "(e.g., contrastive [39, 33, 40, 41, 35] or prediction/reconstruction [18, 34, 37, 36]). For example,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 223, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 237 + ], + "score": 1.0, + "content": "wav2vec 2.0 [33] takes raw waveform as inputs and exploits contrastive learning to discriminate", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 234, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 246 + ], + "score": 1.0, + "content": "contextualized representations in different time segments. Mockingjay [42] proposed a masked", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 245, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 506, + 257 + ], + "score": 1.0, + "content": "acoustic model pretext task to reconstruct frame-level Mel-features of masked time frames. SS-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 257, + 506, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 506, + 268 + ], + "score": 1.0, + "content": "AST [18] is the closest work to Audio-MAE and is our main benchmark. Inspired by the success of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "BERT [3], SS-AST proposed a self-supervised learning method which operates over spectrogram", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 278, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 506, + 291 + ], + "score": 1.0, + "content": "patches and employs joint contrastive and reconstructive objectives on masked patches. These", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "score": 1.0, + "content": "previous methods generate audio representations by encoding full-view of both masked and non-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 299, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 506, + 313 + ], + "score": 1.0, + "content": "masked time or spectrogram segments for self-supervised pre-training. In contrast, Audio-MAE", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 312, + 313, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 313, + 323 + ], + "score": 1.0, + "content": "encodes only the non-masked spectrogram patches.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15, + "bbox_fs": [ + 104, + 179, + 506, + 323 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 327, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "score": 1.0, + "content": "Our work is done independently and concurrently with [38, 43, 44] related methods. We also compare", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 338, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 506, + 350 + ], + "score": 1.0, + "content": "our model to these concurrent works in the experiments and showcase the superiority of Audio-MAE.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 326, + 506, + 350 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 366, + 346, + 380 + ], + "lines": [ + { + "bbox": [ + 104, + 365, + 347, + 383 + ], + "spans": [ + { + "bbox": [ + 104, + 365, + 347, + 383 + ], + "score": 1.0, + "content": "3 Audio Masked Autoencoders (Audio-MAE)", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 504, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "Audio-MAE is a conceptually simple extension of MAE to learn self-supervised representations from", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 403, + 484, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 484, + 416 + ], + "score": 1.0, + "content": "audio spectrograms. Fig. 1 depicts an overview. The details of each component are as follows.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 392, + 505, + 416 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 419, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 419, + 507, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 507, + 432 + ], + "score": 1.0, + "content": "Spectrogram Patch Embeddings. Following [10, 18], we transform audio recordings into Mel-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "score": 1.0, + "content": "spectrograms and divide them into non-overlapped regular grid patches. These patches are then", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 441, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 454 + ], + "score": 1.0, + "content": "flattened and embedded by a linear projection. Similar to MAE [1], we add fixed sinusoidal positional", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 452, + 260, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 260, + 464 + ], + "score": 1.0, + "content": "embeddings to the embedded patches.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 419, + 507, + 464 + ] + }, + { + "type": "image", + "bbox": [ + 112, + 476, + 498, + 543 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 476, + 498, + 543 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 476, + 498, + 543 + ], + "spans": [ + { + "bbox": [ + 112, + 476, + 498, + 543 + ], + "score": 0.969, + "type": "image", + "image_path": "31a937966647bc9c70a9fc197b7e10a2aa38720c7ea2a23229a1c5f045f9bb94.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 112, + 476, + 498, + 498.3333333333333 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 112, + 498.3333333333333, + 498, + 520.6666666666666 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 112, + 520.6666666666666, + 498, + 543.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 172, + 549, + 437, + 561 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 172, + 547, + 439, + 564 + ], + "spans": [ + { + "bbox": [ + 172, + 547, + 439, + 564 + ], + "score": 1.0, + "content": "Figure 2: Audio-MAE’s masking strategies on Mel-spectrograms.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + } + ], + "index": 33.0 + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 505, + 641 + ], + "lines": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "Masking Strategies. Audio-MAE masks out a large subset of spectrogram patches. As a spectrogram", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 585, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 598 + ], + "score": 1.0, + "content": "can be viewed as a 2D representation of time and frequency components of a sound, it is reasonable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "to explore treating time and frequency differently during masking. In this work, we investigate both", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "score": 1.0, + "content": "the unstructured (i.e., random masking without any prior) and structured (i.e., randomly masking a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 617, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 104, + 617, + 245, + 632 + ], + "score": 1.0, + "content": "portion of time, frequency, or time", + "type": "text" + }, + { + "bbox": [ + 245, + 619, + 253, + 628 + ], + "score": 0.33, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 617, + 506, + 632 + ], + "score": 1.0, + "content": "frequency of a spectrogram) in the pre-training and fine-tuning", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 628, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 480, + 642 + ], + "score": 1.0, + "content": "phase. Illustrative examples are shown in Fig. 2. We show masked regions with dark overlay.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5, + "bbox_fs": [ + 104, + 573, + 506, + 642 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "The masking mechanism, as introduced in MAE [1], is the key ingredient for efficient self-supervised", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 656, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 506, + 668 + ], + "score": 1.0, + "content": "learning. For a input patch sequence, this can be regarded as a Bernoulli process where each patch is", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 237, + 680 + ], + "score": 1.0, + "content": "masked/dropped with probability", + "type": "text" + }, + { + "bbox": [ + 238, + 669, + 244, + 679 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "(masking ratio). Masking reduces input patch sequence length and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 679, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 505, + 690 + ], + "score": 1.0, + "content": "encourages learning global, contextualized representations from limited “visible” patches. We observe", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 276, + 701 + ], + "score": 1.0, + "content": "that akin to images, a large masking rate (", + "type": "text" + }, + { + "bbox": [ + 276, + 689, + 296, + 699 + ], + "score": 0.86, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "in our experiments for spectrogram patches, which", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 156, + 713 + ], + "score": 1.0, + "content": "is similar to", + "type": "text" + }, + { + "bbox": [ + 156, + 700, + 176, + 711 + ], + "score": 0.88, + "content": "7 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "in MAE for images) is feasible for learning self-supervised audio representations.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 709, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 222, + 724 + ], + "score": 1.0, + "content": "Unlike BERT [3] that uses", + "type": "text" + }, + { + "bbox": [ + 222, + 711, + 242, + 721 + ], + "score": 0.85, + "content": "15 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 709, + 505, + 724 + ], + "score": 1.0, + "content": "masking rate for self-supervised learning in NLP, most of the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "tokens/patches can be discarded for spectrograms as well as images due to high redundancy in these", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "modalities. Beyond self-supervised pre-training, we further explore the effectiveness of masking in", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 95, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 95, + 505, + 106 + ], + "score": 1.0, + "content": "the supervised fine-tuning stage. Empirically, we found unstructured (random) masking at a higher", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 119 + ], + "score": 1.0, + "content": "ratio for pre-training and structured (time+frequency masking) at a lower ratio for fine-tuning provide", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 114, + 255, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 229, + 131 + ], + "score": 1.0, + "content": "best accuracy (ablations are in", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 230, + 117, + 250, + 128 + ], + "score": 0.32, + "content": "\\ S \\_ 4 )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 250, + 114, + 255, + 131 + ], + "score": 1.0, + "content": ").", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 645, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 73, + 505, + 128 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "tokens/patches can be discarded for spectrograms as well as images due to high redundancy in these", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "modalities. Beyond self-supervised pre-training, we further explore the effectiveness of masking in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 95, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 95, + 505, + 106 + ], + "score": 1.0, + "content": "the supervised fine-tuning stage. Empirically, we found unstructured (random) masking at a higher", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 119 + ], + "score": 1.0, + "content": "ratio for pre-training and structured (time+frequency masking) at a lower ratio for fine-tuning provide", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 114, + 255, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 229, + 131 + ], + "score": 1.0, + "content": "best accuracy (ablations are in", + "type": "text" + }, + { + "bbox": [ + 230, + 117, + 250, + 128 + ], + "score": 0.32, + "content": "\\ S \\_ 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 114, + 255, + 131 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 504, + 166 + ], + "lines": [ + { + "bbox": [ + 106, + 133, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 133, + 505, + 145 + ], + "score": 1.0, + "content": "Encoder. Audio-MAE uses a stack of standard Transformers [2] as its encoder. The encoder only", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 144, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 147, + 156 + ], + "score": 1.0, + "content": "processes", + "type": "text" + }, + { + "bbox": [ + 147, + 144, + 172, + 155 + ], + "score": 0.87, + "content": "( 2 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 144, + 506, + 156 + ], + "score": 1.0, + "content": "non-masked patches to reduce computation overhead which is quadratic to the input", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 456, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 456, + 167 + ], + "score": 1.0, + "content": "sequence length. We use the 12-layer ViT-Base (ViT-B) [9] Transformer as our default.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 171, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 106, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "Decoder with Local Attention. The decoder is also composed of standard Transformer blocks.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 195 + ], + "score": 1.0, + "content": "The encoded patches from the encoder are padded with trainable masked tokens. After restoring", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "the original time-frequency order in the audio spectrogram, we add the decoder’s (fixed sinusoidal)", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "positional embeddings and feed the restored sequence into the decoder. At the top of the decoder", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 214, + 412, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 412, + 228 + ], + "score": 1.0, + "content": "stack, we add a linear head to predict and reconstruct the input spectrogram.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 230, + 505, + 362 + ], + "lines": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "To address the unique characteristics of audio spectrograms, our work investigates an enhancement to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "the vanilla MAE decoder. Image-based MAE uses global self-attention in the Transformer decoder", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "which is appropriate for visual context, because visual objects are typically invariant under translation", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 277 + ], + "score": 1.0, + "content": "or scaling, and their exact position may not affect the semantics of an image. In contrast, the position,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "scale, and translation of spectrogram features however directly affects the sound or semantics of an", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 506, + 298 + ], + "score": 1.0, + "content": "audio recording. Consequently, global self-attention is sub-optimal for spectrograms if the time-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "frequency components is predominantly local. For instance, we would have better success to use the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "score": 1.0, + "content": "harmonics (e.g., Fig. 2a) in lower bands of a vowel to predict the spectrogram patch vertically in a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 318, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 506, + 330 + ], + "score": 1.0, + "content": "higher frequency band rather than horizontally in the time domain. Similarly, a frictional sound of a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 329, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 104, + 329, + 506, + 342 + ], + "score": 1.0, + "content": "consonant likely only correlates to other part of the consonant, and is without dependency to other", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "score": 1.0, + "content": "silence segments in the audio recording. Compared to images, the spectrogram patches are more", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 352, + 415, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 415, + 363 + ], + "score": 1.0, + "content": "similar to speech or text tokens where its order and position is more relevant.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 353, + 476 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 355, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 355, + 379 + ], + "score": 1.0, + "content": "To address the nature of audio spectrograms, in addition to us-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 378, + 354, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 354, + 390 + ], + "score": 1.0, + "content": "ing Transformers with global self-attention as in vanilla MAE,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 388, + 354, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 354, + 402 + ], + "score": 1.0, + "content": "we incorporate the local attention mechanism which groups", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 400, + 354, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 354, + 412 + ], + "score": 1.0, + "content": "and separates the spectrogram patches in to local windows", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 410, + 354, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 354, + 423 + ], + "score": 1.0, + "content": "in self-attention for decoding. We investigate two types of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 421, + 353, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 353, + 434 + ], + "score": 1.0, + "content": "local attention: (1) Shifted window location: Inspired by the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 433, + 354, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 354, + 444 + ], + "score": 1.0, + "content": "shifted-window in Swin Transformers [19], we shift window", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 443, + 354, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 157, + 455 + ], + "score": 1.0, + "content": "attention by", + "type": "text" + }, + { + "bbox": [ + 158, + 444, + 178, + 455 + ], + "score": 0.86, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 443, + 354, + 455 + ], + "score": 1.0, + "content": "between consecutive Transformer decoder", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 455, + 353, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 353, + 467 + ], + "score": 1.0, + "content": "layers. For padding the margin when shifting, we cyclically", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 466, + 354, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 354, + 477 + ], + "score": 1.0, + "content": "shift the spectrogram to the top-left direction. Fig. 3 illustrates", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 30.5 + }, + { + "type": "image", + "bbox": [ + 362, + 377, + 503, + 437 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 362, + 377, + 503, + 437 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 362, + 377, + 503, + 437 + ], + "spans": [ + { + "bbox": [ + 362, + 377, + 503, + 437 + ], + "score": 0.847, + "type": "image", + "image_path": "eac4e377de5817b96d0854d87b3e709eb9a487a2db0842240cefe54d8a061b6c.jpg" + } + ] + } + ], + "index": 30.0, + "virtual_lines": [ + { + "bbox": [ + 362, + 377, + 503, + 407.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 362, + 407.0, + 503, + 437.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 361, + 442, + 504, + 465 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 361, + 442, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 361, + 442, + 505, + 453 + ], + "score": 1.0, + "content": "Figure 3: Decoder’s local attention", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 361, + 453, + 471, + 464 + ], + "spans": [ + { + "bbox": [ + 361, + 453, + 471, + 464 + ], + "score": 1.0, + "content": "and shifted window (right).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.0 + } + ], + "index": 33.0 + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 505, + 542 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "the localized decoder attention by shifted windows. (2) Hybrid window attention (global+local", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "attention): Inspired by [45], to add better cross-window connections, we design a simple hybrid", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 498, + 507, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 507, + 511 + ], + "score": 1.0, + "content": "(global+local) attention that computes local attention within a window in all but the last few top layers.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 510, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 506, + 521 + ], + "score": 1.0, + "content": "In this way, the input feature maps for the final reconstruction layer also contain global information.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 520, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 197, + 533 + ], + "score": 1.0, + "content": "For simplicity, we use", + "type": "text" + }, + { + "bbox": [ + 198, + 522, + 209, + 530 + ], + "score": 0.27, + "content": "_ { n o }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 520, + 506, + 533 + ], + "score": 1.0, + "content": "pooling or hierarchical structure. Decoders with different attention types", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 531, + 195, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 173, + 543 + ], + "score": 1.0, + "content": "are compared in", + "type": "text" + }, + { + "bbox": [ + 173, + 531, + 192, + 542 + ], + "score": 0.83, + "content": "\\ S \\ O = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 531, + 195, + 543 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 546, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 560 + ], + "score": 1.0, + "content": "Objective. The Audio-MAE decoder learns to reconstruct the input spectrogram by predicting the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "values in the spectrogram patches or their per-patch normalized ones. The objective is the mean", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 568, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 583 + ], + "score": 1.0, + "content": "squared error (MSE) between the prediction and the input spectrogram, averaged over unknown", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 579, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 594 + ], + "score": 1.0, + "content": "patches. Empirically we found employing the reconstruction loss alone is sufficient while including", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 591, + 466, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 466, + 604 + ], + "score": 1.0, + "content": "additional contrastive objectives (e.g., InfoNCE loss [46]) does not improve Audio-MAE.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "score": 1.0, + "content": "Fine-tuning for Downstream Tasks. In the fine-tuning stage, we only keep and fine-tune the Audio-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 618, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 631 + ], + "score": 1.0, + "content": "MAE encoder and discard the decoder. Different from the original MAE, and inspired by [47, 28],", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 630, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 506, + 642 + ], + "score": 1.0, + "content": "we also explore to employ masking in the fine-tuning stage to remove a portion of patches to further", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 641, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 505, + 653 + ], + "score": 1.0, + "content": "regularize learning from a limited view of spectrogram inputs, which, as a side effect, also reduces", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "computation during fine-tuning. Compared to SpecAug [48] which takes full-length input with the", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "masked portion set to zero as data augmentation, Audio-MAE sees only a subset of real-valued input", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 672, + 506, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 506, + 686 + ], + "score": 1.0, + "content": "patches without the nullified ones. Audio-MAE then encodes these non-masked patches and applies", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 684, + 490, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 490, + 696 + ], + "score": 1.0, + "content": "an average pooling layer followed by a linear layer on top for fine-tuning in classification tasks.", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 53.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 11, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 73, + 505, + 128 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 105, + 73, + 505, + 131 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 504, + 166 + ], + "lines": [ + { + "bbox": [ + 106, + 133, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 133, + 505, + 145 + ], + "score": 1.0, + "content": "Encoder. Audio-MAE uses a stack of standard Transformers [2] as its encoder. The encoder only", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 144, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 147, + 156 + ], + "score": 1.0, + "content": "processes", + "type": "text" + }, + { + "bbox": [ + 147, + 144, + 172, + 155 + ], + "score": 0.87, + "content": "( 2 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 144, + 506, + 156 + ], + "score": 1.0, + "content": "non-masked patches to reduce computation overhead which is quadratic to the input", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 456, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 456, + 167 + ], + "score": 1.0, + "content": "sequence length. We use the 12-layer ViT-Base (ViT-B) [9] Transformer as our default.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 133, + 506, + 167 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 171, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 106, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "Decoder with Local Attention. The decoder is also composed of standard Transformer blocks.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 195 + ], + "score": 1.0, + "content": "The encoded patches from the encoder are padded with trainable masked tokens. After restoring", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "the original time-frequency order in the audio spectrogram, we add the decoder’s (fixed sinusoidal)", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "positional embeddings and feed the restored sequence into the decoder. At the top of the decoder", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 214, + 412, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 412, + 228 + ], + "score": 1.0, + "content": "stack, we add a linear head to predict and reconstruct the input spectrogram.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 171, + 506, + 228 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 230, + 505, + 362 + ], + "lines": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "To address the unique characteristics of audio spectrograms, our work investigates an enhancement to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "the vanilla MAE decoder. Image-based MAE uses global self-attention in the Transformer decoder", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "which is appropriate for visual context, because visual objects are typically invariant under translation", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 277 + ], + "score": 1.0, + "content": "or scaling, and their exact position may not affect the semantics of an image. In contrast, the position,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "scale, and translation of spectrogram features however directly affects the sound or semantics of an", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 506, + 298 + ], + "score": 1.0, + "content": "audio recording. Consequently, global self-attention is sub-optimal for spectrograms if the time-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "frequency components is predominantly local. For instance, we would have better success to use the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "score": 1.0, + "content": "harmonics (e.g., Fig. 2a) in lower bands of a vowel to predict the spectrogram patch vertically in a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 318, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 506, + 330 + ], + "score": 1.0, + "content": "higher frequency band rather than horizontally in the time domain. Similarly, a frictional sound of a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 329, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 104, + 329, + 506, + 342 + ], + "score": 1.0, + "content": "consonant likely only correlates to other part of the consonant, and is without dependency to other", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "score": 1.0, + "content": "silence segments in the audio recording. Compared to images, the spectrogram patches are more", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 352, + 415, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 415, + 363 + ], + "score": 1.0, + "content": "similar to speech or text tokens where its order and position is more relevant.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 18.5, + "bbox_fs": [ + 104, + 230, + 506, + 363 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 353, + 476 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 355, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 355, + 379 + ], + "score": 1.0, + "content": "To address the nature of audio spectrograms, in addition to us-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 378, + 354, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 354, + 390 + ], + "score": 1.0, + "content": "ing Transformers with global self-attention as in vanilla MAE,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 388, + 354, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 354, + 402 + ], + "score": 1.0, + "content": "we incorporate the local attention mechanism which groups", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 400, + 354, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 354, + 412 + ], + "score": 1.0, + "content": "and separates the spectrogram patches in to local windows", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 410, + 354, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 354, + 423 + ], + "score": 1.0, + "content": "in self-attention for decoding. We investigate two types of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 421, + 353, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 353, + 434 + ], + "score": 1.0, + "content": "local attention: (1) Shifted window location: Inspired by the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 433, + 354, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 354, + 444 + ], + "score": 1.0, + "content": "shifted-window in Swin Transformers [19], we shift window", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 443, + 354, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 157, + 455 + ], + "score": 1.0, + "content": "attention by", + "type": "text" + }, + { + "bbox": [ + 158, + 444, + 178, + 455 + ], + "score": 0.86, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 443, + 354, + 455 + ], + "score": 1.0, + "content": "between consecutive Transformer decoder", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 455, + 353, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 353, + 467 + ], + "score": 1.0, + "content": "layers. For padding the margin when shifting, we cyclically", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 466, + 354, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 354, + 477 + ], + "score": 1.0, + "content": "shift the spectrogram to the top-left direction. Fig. 3 illustrates", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "the localized decoder attention by shifted windows. (2) Hybrid window attention (global+local", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "attention): Inspired by [45], to add better cross-window connections, we design a simple hybrid", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 498, + 507, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 507, + 511 + ], + "score": 1.0, + "content": "(global+local) attention that computes local attention within a window in all but the last few top layers.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 510, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 506, + 521 + ], + "score": 1.0, + "content": "In this way, the input feature maps for the final reconstruction layer also contain global information.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 520, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 197, + 533 + ], + "score": 1.0, + "content": "For simplicity, we use", + "type": "text" + }, + { + "bbox": [ + 198, + 522, + 209, + 530 + ], + "score": 0.27, + "content": "_ { n o }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 520, + 506, + 533 + ], + "score": 1.0, + "content": "pooling or hierarchical structure. Decoders with different attention types", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 531, + 195, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 173, + 543 + ], + "score": 1.0, + "content": "are compared in", + "type": "text" + }, + { + "bbox": [ + 173, + 531, + 192, + 542 + ], + "score": 0.83, + "content": "\\ S \\ O = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 531, + 195, + 543 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 366, + 355, + 477 + ] + }, + { + "type": "image", + "bbox": [ + 362, + 377, + 503, + 437 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 362, + 377, + 503, + 437 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 362, + 377, + 503, + 437 + ], + "spans": [ + { + "bbox": [ + 362, + 377, + 503, + 437 + ], + "score": 0.847, + "type": "image", + "image_path": "eac4e377de5817b96d0854d87b3e709eb9a487a2db0842240cefe54d8a061b6c.jpg" + } + ] + } + ], + "index": 30.0, + "virtual_lines": [ + { + "bbox": [ + 362, + 377, + 503, + 407.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 362, + 407.0, + 503, + 437.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 361, + 442, + 504, + 465 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 361, + 442, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 361, + 442, + 505, + 453 + ], + "score": 1.0, + "content": "Figure 3: Decoder’s local attention", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 361, + 453, + 471, + 464 + ], + "spans": [ + { + "bbox": [ + 361, + 453, + 471, + 464 + ], + "score": 1.0, + "content": "and shifted window (right).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.0 + } + ], + "index": 33.0 + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 505, + 542 + ], + "lines": [], + "index": 41.5, + "bbox_fs": [ + 105, + 476, + 507, + 543 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 546, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 560 + ], + "score": 1.0, + "content": "Objective. The Audio-MAE decoder learns to reconstruct the input spectrogram by predicting the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "values in the spectrogram patches or their per-patch normalized ones. The objective is the mean", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 568, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 583 + ], + "score": 1.0, + "content": "squared error (MSE) between the prediction and the input spectrogram, averaged over unknown", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 579, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 594 + ], + "score": 1.0, + "content": "patches. Empirically we found employing the reconstruction loss alone is sufficient while including", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 591, + 466, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 466, + 604 + ], + "score": 1.0, + "content": "additional contrastive objectives (e.g., InfoNCE loss [46]) does not improve Audio-MAE.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 546, + 506, + 604 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "score": 1.0, + "content": "Fine-tuning for Downstream Tasks. In the fine-tuning stage, we only keep and fine-tune the Audio-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 618, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 631 + ], + "score": 1.0, + "content": "MAE encoder and discard the decoder. Different from the original MAE, and inspired by [47, 28],", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 630, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 506, + 642 + ], + "score": 1.0, + "content": "we also explore to employ masking in the fine-tuning stage to remove a portion of patches to further", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 641, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 505, + 653 + ], + "score": 1.0, + "content": "regularize learning from a limited view of spectrogram inputs, which, as a side effect, also reduces", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "computation during fine-tuning. Compared to SpecAug [48] which takes full-length input with the", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "masked portion set to zero as data augmentation, Audio-MAE sees only a subset of real-valued input", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 672, + 506, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 506, + 686 + ], + "score": 1.0, + "content": "patches without the nullified ones. Audio-MAE then encodes these non-masked patches and applies", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 684, + 490, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 490, + 696 + ], + "score": 1.0, + "content": "an average pooling layer followed by a linear layer on top for fine-tuning in classification tasks.", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 53.5, + "bbox_fs": [ + 105, + 607, + 506, + 696 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 191, + 85 + ], + "lines": [ + { + "bbox": [ + 104, + 69, + 193, + 88 + ], + "spans": [ + { + "bbox": [ + 104, + 69, + 193, + 88 + ], + "score": 1.0, + "content": "4 Experiments", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 96, + 505, + 129 + ], + "lines": [ + { + "bbox": [ + 105, + 96, + 506, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 506, + 109 + ], + "score": 1.0, + "content": "We perform an extensive evaluation on six tasks, including audio classification on AudioSet (AS-2M,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 106, + 505, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 121 + ], + "score": 1.0, + "content": "AS-20K) and Environmental Sound Classification (ESC-50), and speech classification on Speech", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 119, + 474, + 131 + ], + "spans": [ + { + "bbox": [ + 106, + 119, + 474, + 131 + ], + "score": 1.0, + "content": "Commands (SPC-1 and SPC-2) and VoxCeleb (SID). We use AudioSet for ablation studies.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 107, + 144, + 212, + 156 + ], + "lines": [ + { + "bbox": [ + 106, + 144, + 212, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 212, + 156 + ], + "score": 1.0, + "content": "4.1 Datasets and Tasks", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 164, + 505, + 220 + ], + "lines": [ + { + "bbox": [ + 106, + 165, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 282, + 177 + ], + "score": 1.0, + "content": "AudioSet [12] (AS-2M, AS-20K) contains", + "type": "text" + }, + { + "bbox": [ + 282, + 165, + 295, + 175 + ], + "score": 0.78, + "content": "{ \\sim } 2", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 165, + 506, + 177 + ], + "score": 1.0, + "content": "million 10-second YouTube clips for audio classifi-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "cation. 527 types of audio events are weakly annotated [49, 50, 51] for each clip. There could be", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 199 + ], + "score": 1.0, + "content": "multiple events in a clip. The full training set has 2 subsets: A class-wise balanced (22,176 clips) and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "score": 1.0, + "content": "an unbalanced (2,042,985 clips) set. The eval set has 20,383 clips. We downloaded and processed", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 208, + 446, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 446, + 222 + ], + "score": 1.0, + "content": "around 1.96M unbalanced training, 21K balanced training, and 19K evaluation clips.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 225, + 505, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 224, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 506, + 239 + ], + "score": 1.0, + "content": "For the AS-2M experiments, we use the union of unbalanced and balanced training audio for pre-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 235, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 248 + ], + "score": 1.0, + "content": "training and fine-tuning. For the AS-20K experiments, we use AS-2M for pre-training and the 20K", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 246, + 485, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 485, + 259 + ], + "score": 1.0, + "content": "balanced set for fine-tuning. We report the testing mAP on the 19K eval set used by AST [10].", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 505, + 296 + ], + "lines": [ + { + "bbox": [ + 105, + 263, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 275 + ], + "score": 1.0, + "content": "Environmental Sound Classification (ESC-50) [13] is an audio classification dataset consists of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 272, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 505, + 288 + ], + "score": 1.0, + "content": "2,000 5-second environmental sound recordings. There are 50 classes in ESC. We report accuracy", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 284, + 357, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 357, + 298 + ], + "score": 1.0, + "content": "under 5-fold cross-validation with the same split used by [10].", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 301, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "score": 1.0, + "content": "Speech Commands (SPC-2, SPC-1) [52] are two keyword spotting tasks. In SPC-2, there are 35", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 312, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 506, + 325 + ], + "score": 1.0, + "content": "speech commands. The training/validation/testing set has 84,843/9,981/11,005 1-second recordings,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "score": 1.0, + "content": "respectively. In SPC-1, there are 10 classes of keywords, 1 silence class, and 1 unknown class that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 333, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 347 + ], + "score": 1.0, + "content": "includes all the other 20 common speech commands. We use the data and split provided in the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 344, + 332, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 332, + 359 + ], + "score": 1.0, + "content": "SUPERB [53] benchmark to report the testing accuracy.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 394 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 374 + ], + "score": 1.0, + "content": "VoxCeleb (SID) [54] contains 150K utterances from 1,251 speakers. The speaker identification", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "task (SID) is to classify the utterances to identify its original speaker. We use the V1 standard train", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 382, + 431, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 431, + 397 + ], + "score": 1.0, + "content": "(138,361), validation (6,904), testing (8,251) sets and report the testing accuracy.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 108, + 408, + 230, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 231, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 231, + 422 + ], + "score": 1.0, + "content": "4.2 Implementation Details", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 108, + 429, + 504, + 462 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "We use a vanilla 12-layer ViT-B by default as the Transformer encoder. For the decoder, we use a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "16-layer Transformer with shifted local attention. We investigate the vanilla (global attention) and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 452, + 365, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 365, + 464 + ], + "score": 1.0, + "content": "hybrid (global+local attention) decoder variants (see Table. 1c).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 468, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "Following [10, 11], we transform raw waveform (pre-processed as mono channel under 16,000", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 409, + 491 + ], + "score": 1.0, + "content": "sampling rate) into 128 Kaldi [55]-compatible Mel-frequency bands with a", + "type": "text" + }, + { + "bbox": [ + 409, + 479, + 433, + 489 + ], + "score": 0.44, + "content": "2 5 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "Hanning window", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 176, + 502 + ], + "score": 1.0, + "content": "that shifts every", + "type": "text" + }, + { + "bbox": [ + 176, + 490, + 202, + 500 + ], + "score": 0.54, + "content": "1 0 ~ \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 489, + 506, + 502 + ], + "score": 1.0, + "content": ". For a 10-second recording in AudioSet, the resulting spectrogram is of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 499, + 210, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 163, + 511 + ], + "score": 0.89, + "content": "1 \\times 1 0 2 4 \\times 1 2 8", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 499, + 210, + 513 + ], + "score": 1.0, + "content": "dimension.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "For patch embedding, we use convolutional kernels with (16, 16) size and stride in time and frequency", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "(thus, patches are non-overlapping) to avoid short-cuts via overlap in self-supervision (though, at", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "high masking ratios such short-cuts are less severe). By default, we use a masking ratio of 0.8 with", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "(unstructured) random masking for pre-training. During fine-tuning, we employ a lower masking", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 560, + 476, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 453, + 573 + ], + "score": 1.0, + "content": "ratio (0.3 in time and 0.3 in frequency). Ablations on these design choices are given in", + "type": "text" + }, + { + "bbox": [ + 454, + 561, + 473, + 572 + ], + "score": 0.83, + "content": "\\ S \\ O = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 560, + 476, + 573 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + }, + { + "type": "title", + "bbox": [ + 108, + 586, + 254, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 255, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 255, + 601 + ], + "score": 1.0, + "content": "4.3 Pre-training and Fine-tuning", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 607, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 618 + ], + "score": 1.0, + "content": "We use AudioSet-2M for pre-training and randomly iterate over all audio recordings. We train for", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "32 epochs with a batch size of 512 and a 0.0002 learning rate. We distribute the training load over", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 289, + 641 + ], + "score": 1.0, + "content": "64 V100 GPUs and the total training time is", + "type": "text" + }, + { + "bbox": [ + 289, + 629, + 307, + 640 + ], + "score": 0.85, + "content": "{ \\sim } 3 6", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "hours. For each audio, we randomly sample the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 640, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 474, + 651 + ], + "score": 1.0, + "content": "starting time, cyclically extract 10-second audio, and randomly jitter its magnitude by up to", + "type": "text" + }, + { + "bbox": [ + 474, + 640, + 503, + 651 + ], + "score": 0.66, + "content": "\\pm 6 \\mathrm { d B }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 640, + 506, + 651 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 649, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 664 + ], + "score": 1.0, + "content": "We use only natural audio spectrograms and apply no augmentations (e.g., [48, 56, 57]) as we do not", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 661, + 368, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 368, + 675 + ], + "score": 1.0, + "content": "find these strong augmentations helpful in the pre-training phase.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "In the fine-tuning phase, we remove the decoder and only fine-tune the encoder. For the supervised", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "fine-tuning on AudioSet-2M, since the size of training samples are uneven across classes (unbalanced),", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "we follow the common practice of using a weighted sampling to balance the classes during training. In", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 267, + 723 + ], + "score": 1.0, + "content": "each epoch, we sample 200K instances (", + "type": "text" + }, + { + "bbox": [ + 267, + 711, + 293, + 721 + ], + "score": 0.88, + "content": "\\mathord { \\sim } 1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "of AudioSet-2M) without replacement. We fine-tune", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 191, + 85 + ], + "lines": [ + { + "bbox": [ + 104, + 69, + 193, + 88 + ], + "spans": [ + { + "bbox": [ + 104, + 69, + 193, + 88 + ], + "score": 1.0, + "content": "4 Experiments", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 96, + 505, + 129 + ], + "lines": [ + { + "bbox": [ + 105, + 96, + 506, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 506, + 109 + ], + "score": 1.0, + "content": "We perform an extensive evaluation on six tasks, including audio classification on AudioSet (AS-2M,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 106, + 505, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 121 + ], + "score": 1.0, + "content": "AS-20K) and Environmental Sound Classification (ESC-50), and speech classification on Speech", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 119, + 474, + 131 + ], + "spans": [ + { + "bbox": [ + 106, + 119, + 474, + 131 + ], + "score": 1.0, + "content": "Commands (SPC-1 and SPC-2) and VoxCeleb (SID). We use AudioSet for ablation studies.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 96, + 506, + 131 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 144, + 212, + 156 + ], + "lines": [ + { + "bbox": [ + 106, + 144, + 212, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 212, + 156 + ], + "score": 1.0, + "content": "4.1 Datasets and Tasks", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 164, + 505, + 220 + ], + "lines": [ + { + "bbox": [ + 106, + 165, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 282, + 177 + ], + "score": 1.0, + "content": "AudioSet [12] (AS-2M, AS-20K) contains", + "type": "text" + }, + { + "bbox": [ + 282, + 165, + 295, + 175 + ], + "score": 0.78, + "content": "{ \\sim } 2", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 165, + 506, + 177 + ], + "score": 1.0, + "content": "million 10-second YouTube clips for audio classifi-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "cation. 527 types of audio events are weakly annotated [49, 50, 51] for each clip. There could be", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 199 + ], + "score": 1.0, + "content": "multiple events in a clip. The full training set has 2 subsets: A class-wise balanced (22,176 clips) and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "score": 1.0, + "content": "an unbalanced (2,042,985 clips) set. The eval set has 20,383 clips. We downloaded and processed", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 208, + 446, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 446, + 222 + ], + "score": 1.0, + "content": "around 1.96M unbalanced training, 21K balanced training, and 19K evaluation clips.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 165, + 506, + 222 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 225, + 505, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 224, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 506, + 239 + ], + "score": 1.0, + "content": "For the AS-2M experiments, we use the union of unbalanced and balanced training audio for pre-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 235, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 248 + ], + "score": 1.0, + "content": "training and fine-tuning. For the AS-20K experiments, we use AS-2M for pre-training and the 20K", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 246, + 485, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 485, + 259 + ], + "score": 1.0, + "content": "balanced set for fine-tuning. We report the testing mAP on the 19K eval set used by AST [10].", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 224, + 506, + 259 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 505, + 296 + ], + "lines": [ + { + "bbox": [ + 105, + 263, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 275 + ], + "score": 1.0, + "content": "Environmental Sound Classification (ESC-50) [13] is an audio classification dataset consists of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 272, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 505, + 288 + ], + "score": 1.0, + "content": "2,000 5-second environmental sound recordings. There are 50 classes in ESC. We report accuracy", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 284, + 357, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 357, + 298 + ], + "score": 1.0, + "content": "under 5-fold cross-validation with the same split used by [10].", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 263, + 505, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 301, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "score": 1.0, + "content": "Speech Commands (SPC-2, SPC-1) [52] are two keyword spotting tasks. In SPC-2, there are 35", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 312, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 506, + 325 + ], + "score": 1.0, + "content": "speech commands. The training/validation/testing set has 84,843/9,981/11,005 1-second recordings,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "score": 1.0, + "content": "respectively. In SPC-1, there are 10 classes of keywords, 1 silence class, and 1 unknown class that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 333, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 347 + ], + "score": 1.0, + "content": "includes all the other 20 common speech commands. We use the data and split provided in the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 344, + 332, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 332, + 359 + ], + "score": 1.0, + "content": "SUPERB [53] benchmark to report the testing accuracy.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 301, + 506, + 359 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 394 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 374 + ], + "score": 1.0, + "content": "VoxCeleb (SID) [54] contains 150K utterances from 1,251 speakers. The speaker identification", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "task (SID) is to classify the utterances to identify its original speaker. We use the V1 standard train", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 382, + 431, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 431, + 397 + ], + "score": 1.0, + "content": "(138,361), validation (6,904), testing (8,251) sets and report the testing accuracy.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 360, + 505, + 397 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 408, + 230, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 231, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 231, + 422 + ], + "score": 1.0, + "content": "4.2 Implementation Details", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 108, + 429, + 504, + 462 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "We use a vanilla 12-layer ViT-B by default as the Transformer encoder. For the decoder, we use a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "16-layer Transformer with shifted local attention. We investigate the vanilla (global attention) and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 452, + 365, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 365, + 464 + ], + "score": 1.0, + "content": "hybrid (global+local attention) decoder variants (see Table. 1c).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 428, + 506, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 468, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "Following [10, 11], we transform raw waveform (pre-processed as mono channel under 16,000", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 409, + 491 + ], + "score": 1.0, + "content": "sampling rate) into 128 Kaldi [55]-compatible Mel-frequency bands with a", + "type": "text" + }, + { + "bbox": [ + 409, + 479, + 433, + 489 + ], + "score": 0.44, + "content": "2 5 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "Hanning window", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 176, + 502 + ], + "score": 1.0, + "content": "that shifts every", + "type": "text" + }, + { + "bbox": [ + 176, + 490, + 202, + 500 + ], + "score": 0.54, + "content": "1 0 ~ \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 489, + 506, + 502 + ], + "score": 1.0, + "content": ". For a 10-second recording in AudioSet, the resulting spectrogram is of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 499, + 210, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 163, + 511 + ], + "score": 0.89, + "content": "1 \\times 1 0 2 4 \\times 1 2 8", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 499, + 210, + 513 + ], + "score": 1.0, + "content": "dimension.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 467, + 506, + 513 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "For patch embedding, we use convolutional kernels with (16, 16) size and stride in time and frequency", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "(thus, patches are non-overlapping) to avoid short-cuts via overlap in self-supervision (though, at", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "high masking ratios such short-cuts are less severe). By default, we use a masking ratio of 0.8 with", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "(unstructured) random masking for pre-training. During fine-tuning, we employ a lower masking", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 560, + 476, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 453, + 573 + ], + "score": 1.0, + "content": "ratio (0.3 in time and 0.3 in frequency). Ablations on these design choices are given in", + "type": "text" + }, + { + "bbox": [ + 454, + 561, + 473, + 572 + ], + "score": 0.83, + "content": "\\ S \\ O = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 560, + 476, + 573 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 516, + 506, + 573 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 586, + 254, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 255, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 255, + 601 + ], + "score": 1.0, + "content": "4.3 Pre-training and Fine-tuning", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 607, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 618 + ], + "score": 1.0, + "content": "We use AudioSet-2M for pre-training and randomly iterate over all audio recordings. We train for", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "32 epochs with a batch size of 512 and a 0.0002 learning rate. We distribute the training load over", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 289, + 641 + ], + "score": 1.0, + "content": "64 V100 GPUs and the total training time is", + "type": "text" + }, + { + "bbox": [ + 289, + 629, + 307, + 640 + ], + "score": 0.85, + "content": "{ \\sim } 3 6", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "hours. For each audio, we randomly sample the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 640, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 474, + 651 + ], + "score": 1.0, + "content": "starting time, cyclically extract 10-second audio, and randomly jitter its magnitude by up to", + "type": "text" + }, + { + "bbox": [ + 474, + 640, + 503, + 651 + ], + "score": 0.66, + "content": "\\pm 6 \\mathrm { d B }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 640, + 506, + 651 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 649, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 664 + ], + "score": 1.0, + "content": "We use only natural audio spectrograms and apply no augmentations (e.g., [48, 56, 57]) as we do not", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 661, + 368, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 368, + 675 + ], + "score": 1.0, + "content": "find these strong augmentations helpful in the pre-training phase.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 607, + 506, + 675 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "In the fine-tuning phase, we remove the decoder and only fine-tune the encoder. For the supervised", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "fine-tuning on AudioSet-2M, since the size of training samples are uneven across classes (unbalanced),", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "we follow the common practice of using a weighted sampling to balance the classes during training. In", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 267, + 723 + ], + "score": 1.0, + "content": "each epoch, we sample 200K instances (", + "type": "text" + }, + { + "bbox": [ + 267, + 711, + 293, + 721 + ], + "score": 0.88, + "content": "\\mathord { \\sim } 1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "of AudioSet-2M) without replacement. We fine-tune", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 209, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 251, + 224 + ], + "score": 1.0, + "content": "for 100 epochs, which aggregate to", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 251, + 210, + 269, + 221 + ], + "score": 0.85, + "content": "{ \\sim } 1 0", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 270, + 209, + 505, + 224 + ], + "score": 1.0, + "content": "full epochs of AudioSet-2M. The probability of sampling", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 222, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 234 + ], + "score": 1.0, + "content": "an instance is inversely proportional to the dataset-wise occurrences of its classes. Fine-tuning on", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 232, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 168, + 244 + ], + "score": 1.0, + "content": "64 GPUs takes", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 168, + 232, + 186, + 243 + ], + "score": 0.83, + "content": "{ \\sim } 1 2", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 186, + 232, + 505, + 244 + ], + "score": 1.0, + "content": "hours. For the smaller balanced AudioSet-20K, we fine-tune on 4 GPUs for 60", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 244, + 486, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 486, + 255 + ], + "score": 1.0, + "content": "epochs without weighted sampling. Please see Supplementary for the details on other datasets.", + "type": "text", + "cross_page": true + } + ], + "index": 9 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 678, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 69, + 497, + 157 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 69, + 497, + 157 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 69, + 497, + 157 + ], + "spans": [ + { + "bbox": [ + 111, + 69, + 497, + 157 + ], + "score": 0.891, + "type": "image", + "image_path": "77981a7649efa2be98251d63c87ce0bb29948d0455d85ac60829270b7d834094.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 69, + 497, + 98.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 98.33333333333333, + 497, + 127.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 127.66666666666666, + 497, + 157.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 163, + 505, + 196 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 163, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 506, + 176 + ], + "score": 1.0, + "content": "Figure 4: Masking strategy. For pre-training, a higher ratio and unstructured masking (random)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 174, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 397, + 186 + ], + "score": 1.0, + "content": "is preferred. For fine-tuning, a lower ratio and structured masking (time", + "type": "text" + }, + { + "bbox": [ + 397, + 175, + 404, + 185 + ], + "score": 0.39, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 174, + 505, + 186 + ], + "score": 1.0, + "content": "frequency) is better. The", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 185, + 484, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 248, + 198 + ], + "score": 1.0, + "content": "y-axes are mAP on AS-2M and the", + "type": "text" + }, + { + "bbox": [ + 248, + 187, + 255, + 195 + ], + "score": 0.31, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 185, + 484, + 198 + ], + "score": 1.0, + "content": "-axes are masking ratio. This ablation format follows [1].", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 210, + 505, + 254 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 251, + 224 + ], + "score": 1.0, + "content": "for 100 epochs, which aggregate to", + "type": "text" + }, + { + "bbox": [ + 251, + 210, + 269, + 221 + ], + "score": 0.85, + "content": "{ \\sim } 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 209, + 505, + 224 + ], + "score": 1.0, + "content": "full epochs of AudioSet-2M. The probability of sampling", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 222, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 234 + ], + "score": 1.0, + "content": "an instance is inversely proportional to the dataset-wise occurrences of its classes. Fine-tuning on", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 232, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 168, + 244 + ], + "score": 1.0, + "content": "64 GPUs takes", + "type": "text" + }, + { + "bbox": [ + 168, + 232, + 186, + 243 + ], + "score": 0.83, + "content": "{ \\sim } 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 232, + 505, + 244 + ], + "score": 1.0, + "content": "hours. For the smaller balanced AudioSet-20K, we fine-tune on 4 GPUs for 60", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 244, + 486, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 486, + 255 + ], + "score": 1.0, + "content": "epochs without weighted sampling. Please see Supplementary for the details on other datasets.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "title", + "bbox": [ + 108, + 270, + 266, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 267, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 267, + 283 + ], + "score": 1.0, + "content": "4.4 Ablations and Model Properties", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 505, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 293, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 308 + ], + "score": 1.0, + "content": "Masking Strategies in Pre-training and Fine-tuning. In Fig. 4, we compare different pre-training", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 304, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 319 + ], + "score": 1.0, + "content": "and fine-tuning masking strategies for Audio-MAE. First, in Fig. 4a we explore the pre-training", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "score": 1.0, + "content": "masking ratio. We observe, similar as in MAE for images [1], that a high pre-training masking", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 327, + 504, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 131, + 339 + ], + "score": 1.0, + "content": "ratio", + "type": "text" + }, + { + "bbox": [ + 132, + 327, + 151, + 338 + ], + "score": 0.85, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 327, + 504, + 339 + ], + "score": 1.0, + "content": "in our case) is optimal for audio spectrograms. This is due to the fact that both audio", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 339, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 505, + 350 + ], + "score": 1.0, + "content": "spectrograms and images are continuous signals with significant redundancy. Further, we find the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 349, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 362 + ], + "score": 1.0, + "content": "unstructured random masking works the best for self-supervised pre-training over more structured", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 360, + 239, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 239, + 373 + ], + "score": 1.0, + "content": "masking (e.g., time+frequency).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 376, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "Unlike MAE for images, there are clear performance differences among masking strategies when", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 386, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 401 + ], + "score": 1.0, + "content": "pre-training with audio spectrograms. Comparing Audio-MAE reconstructions between Fig. 6a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "score": 1.0, + "content": "to 6e and 6d to 6h, under the same masking ratio, we observe the unstructured random masking", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 409, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 505, + 422 + ], + "score": 1.0, + "content": "is comparably easier than structured masking (i.e., time and/or frequency) as the model can guess", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "the missing component by extrapolating nearby context (e.g., formants in vowels and frictional", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 430, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 505, + 444 + ], + "score": 1.0, + "content": "sounds in consonants around). We also observe that for higher masking ratios, the structured masking", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 442, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 454 + ], + "score": 1.0, + "content": "alternatives drop in performance, presumably because the task becomes too difficult while random", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 241, + 466 + ], + "score": 1.0, + "content": "masking improves steadily up to", + "type": "text" + }, + { + "bbox": [ + 242, + 453, + 262, + 463 + ], + "score": 0.88, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 452, + 505, + 466 + ], + "score": 1.0, + "content": ". This result show that designing a pretext task with proper", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "score": 1.0, + "content": "hardness is important for effective self-supervised learning of audio representations. We therefore", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 474, + 387, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 243, + 489 + ], + "score": 1.0, + "content": "use random masking with ratio of", + "type": "text" + }, + { + "bbox": [ + 243, + 475, + 263, + 485 + ], + "score": 0.87, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 474, + 387, + 489 + ], + "score": 1.0, + "content": "as our default for pre-training.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 491, + 505, + 535 + ], + "lines": [ + { + "bbox": [ + 105, + 491, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 504 + ], + "score": 1.0, + "content": "Fig. 4b studies the effect of masking during the fine-tuning phase. We see that in this case, it is more", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "score": 1.0, + "content": "beneficial to use structured masking: time+frequency performs better than time- or frequency-based", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 513, + 506, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 526 + ], + "score": 1.0, + "content": "masking, and these perform better than unstructured masking. Overall, we see that the optimal", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 524, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 506, + 536 + ], + "score": 1.0, + "content": "masking ratios are lower than for pre-training and we use 0.3 as our default in the fine-tuning phase.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 540, + 504, + 562 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 553 + ], + "score": 1.0, + "content": "In general, we observe that for task-agnostic pre-training, unstructured masking with a higher ratio is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 551, + 501, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 501, + 564 + ], + "score": 1.0, + "content": "preferred. While in task-specific fine-tuning, structured masking with lower ratios performs better.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 570, + 505, + 669 + ], + "lines": [ + { + "bbox": [ + 106, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "Impact of Patch Size and Stride. We compare the performance of Audio-MAE trained with different", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 366, + 595 + ], + "score": 1.0, + "content": "patch sizes and strides in Table 1a. A non-zero overlap (i.e., stride", + "type": "text" + }, + { + "bbox": [ + 366, + 583, + 376, + 592 + ], + "score": 0.75, + "content": "<", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "patch size) between patches will", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 592, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 606 + ], + "score": 1.0, + "content": "increase the number of patches and quadratically increase computation in floating point operations", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "score": 1.0, + "content": "(FLOPs), as reported in the table. Most prior works follow AST [10] to use overlapped patches", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 614, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 132, + 628 + ], + "score": 1.0, + "content": "(patch", + "type": "text" + }, + { + "bbox": [ + 132, + 615, + 155, + 625 + ], + "score": 0.74, + "content": "= 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 614, + 195, + 628 + ], + "score": 1.0, + "content": "and stride", + "type": "text" + }, + { + "bbox": [ + 195, + 615, + 217, + 625 + ], + "score": 0.84, + "content": "= 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 614, + 506, + 628 + ], + "score": 1.0, + "content": ") to boost end task performance. As shown in Table 1a, we do not observe", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 625, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 409, + 638 + ], + "score": 1.0, + "content": "a performance improvement using overlapped patches for Audio-MAE (both", + "type": "text" + }, + { + "bbox": [ + 409, + 626, + 451, + 636 + ], + "score": 0.45, + "content": "4 7 . 3 \\mathrm { m A P }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 625, + 505, + 638 + ], + "score": 1.0, + "content": "), presumably", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "because due to overlap, the patch embedding can leak information into the masked patches. The", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 172, + 660 + ], + "score": 1.0, + "content": "non-overlapped", + "type": "text" + }, + { + "bbox": [ + 172, + 647, + 201, + 658 + ], + "score": 0.89, + "content": "1 6 \\times 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 646, + 505, + 660 + ], + "score": 1.0, + "content": "patches achieve a good balance between computation and performance. By", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 657, + 288, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 288, + 672 + ], + "score": 1.0, + "content": "default, we use this setup in our experiments.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "Encoder. We investigate the design choices of encoder and decoder architectures in Audio-MAE.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "Table 1b shows the trade-off between encoder model size and performance. As expected, larger", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 700, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 507, + 713 + ], + "score": 1.0, + "content": "models achieve better performance, at a cost of computation and memory. The accuracy gain of ViT-L", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "score": 1.0, + "content": "over ViT-B/S is more significant on the smaller and balanced AS-20K. For ViT-S, the performance", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 69, + 497, + 157 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 69, + 497, + 157 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 69, + 497, + 157 + ], + "spans": [ + { + "bbox": [ + 111, + 69, + 497, + 157 + ], + "score": 0.891, + "type": "image", + "image_path": "77981a7649efa2be98251d63c87ce0bb29948d0455d85ac60829270b7d834094.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 69, + 497, + 98.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 98.33333333333333, + 497, + 127.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 127.66666666666666, + 497, + 157.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 163, + 505, + 196 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 163, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 506, + 176 + ], + "score": 1.0, + "content": "Figure 4: Masking strategy. For pre-training, a higher ratio and unstructured masking (random)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 174, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 397, + 186 + ], + "score": 1.0, + "content": "is preferred. For fine-tuning, a lower ratio and structured masking (time", + "type": "text" + }, + { + "bbox": [ + 397, + 175, + 404, + 185 + ], + "score": 0.39, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 174, + 505, + 186 + ], + "score": 1.0, + "content": "frequency) is better. The", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 185, + 484, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 248, + 198 + ], + "score": 1.0, + "content": "y-axes are mAP on AS-2M and the", + "type": "text" + }, + { + "bbox": [ + 248, + 187, + 255, + 195 + ], + "score": 0.31, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 185, + 484, + 198 + ], + "score": 1.0, + "content": "-axes are masking ratio. This ablation format follows [1].", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 210, + 505, + 254 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 105, + 209, + 505, + 255 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 270, + 266, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 267, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 267, + 283 + ], + "score": 1.0, + "content": "4.4 Ablations and Model Properties", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 505, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 293, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 308 + ], + "score": 1.0, + "content": "Masking Strategies in Pre-training and Fine-tuning. In Fig. 4, we compare different pre-training", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 304, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 319 + ], + "score": 1.0, + "content": "and fine-tuning masking strategies for Audio-MAE. First, in Fig. 4a we explore the pre-training", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "score": 1.0, + "content": "masking ratio. We observe, similar as in MAE for images [1], that a high pre-training masking", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 327, + 504, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 131, + 339 + ], + "score": 1.0, + "content": "ratio", + "type": "text" + }, + { + "bbox": [ + 132, + 327, + 151, + 338 + ], + "score": 0.85, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 327, + 504, + 339 + ], + "score": 1.0, + "content": "in our case) is optimal for audio spectrograms. This is due to the fact that both audio", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 339, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 505, + 350 + ], + "score": 1.0, + "content": "spectrograms and images are continuous signals with significant redundancy. Further, we find the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 349, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 362 + ], + "score": 1.0, + "content": "unstructured random masking works the best for self-supervised pre-training over more structured", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 360, + 239, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 239, + 373 + ], + "score": 1.0, + "content": "masking (e.g., time+frequency).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 293, + 506, + 373 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 376, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "Unlike MAE for images, there are clear performance differences among masking strategies when", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 386, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 401 + ], + "score": 1.0, + "content": "pre-training with audio spectrograms. Comparing Audio-MAE reconstructions between Fig. 6a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "score": 1.0, + "content": "to 6e and 6d to 6h, under the same masking ratio, we observe the unstructured random masking", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 409, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 505, + 422 + ], + "score": 1.0, + "content": "is comparably easier than structured masking (i.e., time and/or frequency) as the model can guess", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "the missing component by extrapolating nearby context (e.g., formants in vowels and frictional", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 430, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 505, + 444 + ], + "score": 1.0, + "content": "sounds in consonants around). We also observe that for higher masking ratios, the structured masking", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 442, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 454 + ], + "score": 1.0, + "content": "alternatives drop in performance, presumably because the task becomes too difficult while random", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 241, + 466 + ], + "score": 1.0, + "content": "masking improves steadily up to", + "type": "text" + }, + { + "bbox": [ + 242, + 453, + 262, + 463 + ], + "score": 0.88, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 452, + 505, + 466 + ], + "score": 1.0, + "content": ". This result show that designing a pretext task with proper", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "score": 1.0, + "content": "hardness is important for effective self-supervised learning of audio representations. We therefore", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 474, + 387, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 243, + 489 + ], + "score": 1.0, + "content": "use random masking with ratio of", + "type": "text" + }, + { + "bbox": [ + 243, + 475, + 263, + 485 + ], + "score": 0.87, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 474, + 387, + 489 + ], + "score": 1.0, + "content": "as our default for pre-training.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 375, + 506, + 489 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 491, + 505, + 535 + ], + "lines": [ + { + "bbox": [ + 105, + 491, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 504 + ], + "score": 1.0, + "content": "Fig. 4b studies the effect of masking during the fine-tuning phase. We see that in this case, it is more", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "score": 1.0, + "content": "beneficial to use structured masking: time+frequency performs better than time- or frequency-based", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 513, + 506, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 526 + ], + "score": 1.0, + "content": "masking, and these perform better than unstructured masking. Overall, we see that the optimal", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 524, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 506, + 536 + ], + "score": 1.0, + "content": "masking ratios are lower than for pre-training and we use 0.3 as our default in the fine-tuning phase.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 491, + 506, + 536 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 540, + 504, + 562 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 553 + ], + "score": 1.0, + "content": "In general, we observe that for task-agnostic pre-training, unstructured masking with a higher ratio is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 551, + 501, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 501, + 564 + ], + "score": 1.0, + "content": "preferred. While in task-specific fine-tuning, structured masking with lower ratios performs better.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 540, + 505, + 564 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 570, + 505, + 669 + ], + "lines": [ + { + "bbox": [ + 106, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "Impact of Patch Size and Stride. We compare the performance of Audio-MAE trained with different", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 366, + 595 + ], + "score": 1.0, + "content": "patch sizes and strides in Table 1a. A non-zero overlap (i.e., stride", + "type": "text" + }, + { + "bbox": [ + 366, + 583, + 376, + 592 + ], + "score": 0.75, + "content": "<", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "patch size) between patches will", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 592, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 606 + ], + "score": 1.0, + "content": "increase the number of patches and quadratically increase computation in floating point operations", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "score": 1.0, + "content": "(FLOPs), as reported in the table. Most prior works follow AST [10] to use overlapped patches", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 614, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 132, + 628 + ], + "score": 1.0, + "content": "(patch", + "type": "text" + }, + { + "bbox": [ + 132, + 615, + 155, + 625 + ], + "score": 0.74, + "content": "= 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 614, + 195, + 628 + ], + "score": 1.0, + "content": "and stride", + "type": "text" + }, + { + "bbox": [ + 195, + 615, + 217, + 625 + ], + "score": 0.84, + "content": "= 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 614, + 506, + 628 + ], + "score": 1.0, + "content": ") to boost end task performance. As shown in Table 1a, we do not observe", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 625, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 409, + 638 + ], + "score": 1.0, + "content": "a performance improvement using overlapped patches for Audio-MAE (both", + "type": "text" + }, + { + "bbox": [ + 409, + 626, + 451, + 636 + ], + "score": 0.45, + "content": "4 7 . 3 \\mathrm { m A P }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 625, + 505, + 638 + ], + "score": 1.0, + "content": "), presumably", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "because due to overlap, the patch embedding can leak information into the masked patches. The", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 172, + 660 + ], + "score": 1.0, + "content": "non-overlapped", + "type": "text" + }, + { + "bbox": [ + 172, + 647, + 201, + 658 + ], + "score": 0.89, + "content": "1 6 \\times 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 646, + 505, + 660 + ], + "score": 1.0, + "content": "patches achieve a good balance between computation and performance. By", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 657, + 288, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 288, + 672 + ], + "score": 1.0, + "content": "default, we use this setup in our experiments.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 571, + 506, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "Encoder. We investigate the design choices of encoder and decoder architectures in Audio-MAE.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "Table 1b shows the trade-off between encoder model size and performance. As expected, larger", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 700, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 507, + 713 + ], + "score": 1.0, + "content": "models achieve better performance, at a cost of computation and memory. The accuracy gain of ViT-L", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "score": 1.0, + "content": "over ViT-B/S is more significant on the smaller and balanced AS-20K. For ViT-S, the performance", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 678, + 507, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 76, + 281, + 132 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 123, + 76, + 281, + 132 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 123, + 76, + 281, + 132 + ], + "spans": [ + { + "bbox": [ + 123, + 76, + 281, + 132 + ], + "score": 0.246, + "html": "
(16,16), (16,16)64×848.647.3
(16,16), (10,10)101×12130.547.3
(32,16), (16,16)63×847.846.6
(16,32), (16,16)64×742.146.8
", + "type": "table", + "image_path": "6f59a1c28e76f7aa6d92f54dc2336e032fe1fec05ebb7a82d7a9a980d84c69bf.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 123, + 76, + 281, + 90.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 123, + 90.0, + 281, + 104.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 123, + 104.0, + 281, + 118.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 123, + 118.0, + 281, + 132.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.5 + }, + { + "type": "table", + "bbox": [ + 330, + 82, + 464, + 117 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 330, + 82, + 464, + 117 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 330, + 82, + 464, + 117 + ], + "spans": [ + { + "bbox": [ + 330, + 82, + 464, + 117 + ], + "score": 0.493, + "html": "
ViT-S22M32.145.0
ViT-B86M37.147.3
ViT-L304M37.647.4
", + "type": "table", + "image_path": "cd323607a9ef4d1435f53f5a4694198cfcaa19480ab62f7674c6087d91c726d9.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 330, + 82, + 464, + 99.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 330, + 99.5, + 464, + 117.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 4.5 + }, + { + "type": "table", + "bbox": [ + 118, + 149, + 330, + 199 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 118, + 149, + 330, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 149, + 330, + 199 + ], + "spans": [ + { + "bbox": [ + 118, + 149, + 330, + 199 + ], + "score": 0.154, + "html": "
Attention typeAS-20K AS-2M ESC-50 SID
Global(8) (vanilla)36.6 46.8
93.6 94.1 47.3 94.1Local(16) (shifted) 37.1
Hwin (local(8)+ global(4) 36.894.8 47.3 93.8 95.0
", + "type": "table", + "image_path": "2e64e580b9f04e22b287bd69d5bfdd39254f01c2e6f243ca7c94c7792437d8fd.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 118, + 149, + 330, + 161.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 118, + 161.5, + 330, + 174.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 118, + 174.0, + 330, + 186.5 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 118, + 186.5, + 330, + 199.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 312, + 219, + 504, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 309, + 218, + 504, + 231 + ], + "spans": [ + { + "bbox": [ + 309, + 218, + 504, + 231 + ], + "score": 1.0, + "content": "scenario IN-SSL IN-SL AS-SSL AS-20K AS-2M", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 8.75 + }, + { + "type": "image", + "bbox": [ + 308, + 231, + 507, + 288 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 308, + 231, + 507, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 507, + 288 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 507, + 288 + ], + "score": 0.553, + "type": "image", + "image_path": "c8970b78f179014b4562bee38cb0d12fbbb43623ba5e47ab3c3b6f7e874f2872.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 308, + 231, + 507, + 245.25 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 308, + 245.25, + 507, + 259.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 308, + 259.5, + 507, + 273.75 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 308, + 273.75, + 507, + 288.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 302, + 290, + 509, + 310 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 301, + 289, + 510, + 300 + ], + "spans": [ + { + "bbox": [ + 301, + 289, + 510, + 300 + ], + "score": 1.0, + "content": "(h) External ImageNet (IN) pre-training. SSL: w/ self-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 300, + 299, + 502, + 311 + ], + "spans": [ + { + "bbox": [ + 300, + 299, + 502, + 311 + ], + "score": 1.0, + "content": "supervised MAE. SL: w/ supervised (fine-tuned) MAE.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + } + ], + "index": 14.0 + }, + { + "type": "text", + "bbox": [ + 107, + 314, + 505, + 339 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 504, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 504, + 327 + ], + "score": 1.0, + "content": "Table 1: Ablation studies on AS-2M. The gray entries are the default Audio-MAE setup (ViT-B", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 326, + 500, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 500, + 340 + ], + "score": 1.0, + "content": "encoder, decoder with shifted local attention, pre-trained for 32 epochs). Table format follows [1].", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 504, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 350, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 274, + 365 + ], + "score": 1.0, + "content": "gap to ViT-B can be significantly closed", + "type": "text" + }, + { + "bbox": [ + 275, + 351, + 342, + 362 + ], + "score": 0.77, + "content": "\\mathrm { 5 . 0 \\to 2 . 3 \\ m A P }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 350, + 506, + 365 + ], + "score": 1.0, + "content": ") when fine-tuning with more in-domain", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 361, + 216, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 128, + 374 + ], + "score": 1.0, + "content": "data", + "type": "text" + }, + { + "bbox": [ + 128, + 362, + 210, + 373 + ], + "score": 0.76, + "content": "( \\mathrm { A S } - 2 0 \\mathrm { K } \\mathrm { A S } - 2 \\mathrm { M } )", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 361, + 216, + 374 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 382, + 325, + 523 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 325, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 325, + 394 + ], + "score": 1.0, + "content": "Decoder. Table 1c compares decoder attention types in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 392, + 326, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 326, + 405 + ], + "score": 1.0, + "content": "Audio-MAE. Note that decoders are discarded after pre-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 404, + 325, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 325, + 415 + ], + "score": 1.0, + "content": "training and only the equal-sized ViT-B encoders are", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 415, + 325, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 325, + 425 + ], + "score": 1.0, + "content": "fine-tuned for the end task. Our results show that local", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 425, + 326, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 326, + 437 + ], + "score": 1.0, + "content": "attention with shifted window achieves the best per-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 436, + 326, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 326, + 448 + ], + "score": 1.0, + "content": "formance. Combining local and global attention (i.e.,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 447, + 326, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 326, + 459 + ], + "score": 1.0, + "content": "hybrid attention, Hwin) also improves vanilla global", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 458, + 326, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 326, + 470 + ], + "score": 1.0, + "content": "self-attention. Fig. 5 shows the qualitative reconstruc-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 470, + 326, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 326, + 481 + ], + "score": 1.0, + "content": "tion comparison. In the spectrogram of vowels, the de-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 481, + 325, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 325, + 491 + ], + "score": 1.0, + "content": "coder with local attention reconstructs better harmonics", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 491, + 325, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 325, + 502 + ], + "score": 1.0, + "content": "and recovers more context in the spectrogram. Similar", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 502, + 325, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 325, + 513 + ], + "score": 1.0, + "content": "phenomena are observed in the frictional sound in the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 513, + 182, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 182, + 524 + ], + "score": 1.0, + "content": "middle consonant.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 333, + 395, + 504, + 481 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 333, + 395, + 504, + 481 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 333, + 395, + 504, + 481 + ], + "spans": [ + { + "bbox": [ + 333, + 395, + 504, + 481 + ], + "score": 0.922, + "type": "image", + "image_path": "2eac13304af84b06c5760b4abdf221924cb45af9f00f2bcb95dca8f6af1aa8b1.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 333, + 395, + 504, + 407.2857142857143 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 333, + 407.2857142857143, + 504, + 419.57142857142856 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 333, + 419.57142857142856, + 504, + 431.85714285714283 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 333, + 431.85714285714283, + 504, + 444.1428571428571 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 333, + 444.1428571428571, + 504, + 456.4285714285714 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 333, + 456.4285714285714, + 504, + 468.71428571428567 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 333, + 468.71428571428567, + 504, + 480.99999999999994 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 334, + 487, + 501, + 497 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 332, + 486, + 501, + 499 + ], + "spans": [ + { + "bbox": [ + 332, + 486, + 501, + 499 + ], + "score": 1.0, + "content": "Figure 5: Decoder reconstruction comparison.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + } + ], + "index": 39.0 + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "Table 1d ablates the impact of decoder depth on mAP. A deeper 16-layer decoder achieves better", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "performance against its shallower variants. Note that our decoder uses local window attention by", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 273, + 564 + ], + "score": 1.0, + "content": "default where only a fraction of tokens", + "type": "text" + }, + { + "bbox": [ + 273, + 551, + 293, + 561 + ], + "score": 0.86, + "content": "4 { \\times } 4", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 550, + 372, + 564 + ], + "score": 1.0, + "content": "local windows vs.", + "type": "text" + }, + { + "bbox": [ + 372, + 551, + 397, + 561 + ], + "score": 0.89, + "content": "6 4 \\times 8", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 550, + 506, + 564 + ], + "score": 1.0, + "content": "with global attention) are", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "attended. For global attention we find 8-layer decoders to perform better than 16-layer. Table 1e", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 104, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "compares decoder width (embedding dimension). A 512-dimension decoder achieves a good trade-off", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 584, + 375, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 375, + 595 + ], + "score": 1.0, + "content": "between computation and performance as a wider one is not better.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 107, + 603, + 505, + 669 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 615 + ], + "score": 1.0, + "content": "Pre-training Data and Setup. Table 1f summarizes the impact of pre-training dataset size. Overall", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "score": 1.0, + "content": "the model performance is monotonically increasing when using more data for pre-training. Comparing", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 625, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 208, + 637 + ], + "score": 1.0, + "content": "the performance of using", + "type": "text" + }, + { + "bbox": [ + 208, + 626, + 222, + 636 + ], + "score": 0.85, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 625, + 505, + 637 + ], + "score": 1.0, + "content": "well-annotated AS-20K balanced data to using randomly sampled 20K", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 636, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 648 + ], + "score": 1.0, + "content": "unbalanced data for pre-training, the similar mAPs (39.4 vs 39.6) suggest that the distribution of", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "data classes (balanced vs. unbalanced) is less important for pre-training. Meanwhile, as shown in", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 658, + 478, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 131, + 671 + ], + "score": 1.0, + "content": "Table", + "type": "text" + }, + { + "bbox": [ + 131, + 659, + 142, + 670 + ], + "score": 0.27, + "content": "1 \\mathrm { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 658, + 478, + 671 + ], + "score": 1.0, + "content": ", training for longer is beneficial yet the performance saturates after the 24-th epoch.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50.5 + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 691 + ], + "score": 1.0, + "content": "Out-of-domain Pre-training on ImageNet. Initializing audio models from ImageNet pre-trained", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "weights has become popular for audio classification. However, as there are significant discrepancies", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "between image and audio modalities, it is questionable if out-of-domain pre-training benefits audio", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "representation learning. In Table 1h we design 3 scenarios to investigate this for Audio-MAE: (1)", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 55.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 331, + 71, + 463, + 80 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 463, + 82 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 463, + 82 + ], + "score": 1.0, + "content": "Backbone #Params AS-20K AS-2M", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 125, + 71, + 280, + 81 + ], + "lines": [ + { + "bbox": [ + 125, + 70, + 280, + 82 + ], + "spans": [ + { + "bbox": [ + 125, + 70, + 280, + 82 + ], + "score": 1.0, + "content": "Patch size, stride Seq shape FLOPs mAP", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 76, + 281, + 132 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 123, + 76, + 281, + 132 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 123, + 76, + 281, + 132 + ], + "spans": [ + { + "bbox": [ + 123, + 76, + 281, + 132 + ], + "score": 0.246, + "html": "
(16,16), (16,16)64×848.647.3
(16,16), (10,10)101×12130.547.3
(32,16), (16,16)63×847.846.6
(16,32), (16,16)64×742.146.8
", + "type": "table", + "image_path": "6f59a1c28e76f7aa6d92f54dc2336e032fe1fec05ebb7a82d7a9a980d84c69bf.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 123, + 76, + 281, + 90.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 123, + 90.0, + 281, + 104.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 123, + 104.0, + 281, + 118.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 123, + 118.0, + 281, + 132.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.5 + }, + { + "type": "table", + "bbox": [ + 330, + 82, + 464, + 117 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 330, + 82, + 464, + 117 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 330, + 82, + 464, + 117 + ], + "spans": [ + { + "bbox": [ + 330, + 82, + 464, + 117 + ], + "score": 0.493, + "html": "
ViT-S22M32.145.0
ViT-B86M37.147.3
ViT-L304M37.647.4
", + "type": "table", + "image_path": "cd323607a9ef4d1435f53f5a4694198cfcaa19480ab62f7674c6087d91c726d9.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 330, + 82, + 464, + 99.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 330, + 99.5, + 464, + 117.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 4.5 + }, + { + "type": "table", + "bbox": [ + 118, + 149, + 330, + 199 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 118, + 149, + 330, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 149, + 330, + 199 + ], + "spans": [ + { + "bbox": [ + 118, + 149, + 330, + 199 + ], + "score": 0.154, + "html": "
Attention typeAS-20K AS-2M ESC-50 SID
Global(8) (vanilla)36.6 46.8
93.6 94.1 47.3 94.1Local(16) (shifted) 37.1
Hwin (local(8)+ global(4) 36.894.8 47.3 93.8 95.0
", + "type": "table", + "image_path": "2e64e580b9f04e22b287bd69d5bfdd39254f01c2e6f243ca7c94c7792437d8fd.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 118, + 149, + 330, + 161.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 118, + 161.5, + 330, + 174.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 118, + 174.0, + 330, + 186.5 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 118, + 186.5, + 330, + 199.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 312, + 219, + 504, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 309, + 218, + 504, + 231 + ], + "spans": [ + { + "bbox": [ + 309, + 218, + 504, + 231 + ], + "score": 1.0, + "content": "scenario IN-SSL IN-SL AS-SSL AS-20K AS-2M", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 8.75 + }, + { + "type": "image", + "bbox": [ + 308, + 231, + 507, + 288 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 308, + 231, + 507, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 507, + 288 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 507, + 288 + ], + "score": 0.553, + "type": "image", + "image_path": "c8970b78f179014b4562bee38cb0d12fbbb43623ba5e47ab3c3b6f7e874f2872.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 308, + 231, + 507, + 245.25 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 308, + 245.25, + 507, + 259.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 308, + 259.5, + 507, + 273.75 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 308, + 273.75, + 507, + 288.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 302, + 290, + 509, + 310 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 301, + 289, + 510, + 300 + ], + "spans": [ + { + "bbox": [ + 301, + 289, + 510, + 300 + ], + "score": 1.0, + "content": "(h) External ImageNet (IN) pre-training. SSL: w/ self-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 300, + 299, + 502, + 311 + ], + "spans": [ + { + "bbox": [ + 300, + 299, + 502, + 311 + ], + "score": 1.0, + "content": "supervised MAE. SL: w/ supervised (fine-tuned) MAE.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + } + ], + "index": 14.0 + }, + { + "type": "text", + "bbox": [ + 107, + 314, + 505, + 339 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 504, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 504, + 327 + ], + "score": 1.0, + "content": "Table 1: Ablation studies on AS-2M. The gray entries are the default Audio-MAE setup (ViT-B", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 326, + 500, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 500, + 340 + ], + "score": 1.0, + "content": "encoder, decoder with shifted local attention, pre-trained for 32 epochs). Table format follows [1].", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 313, + 504, + 340 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 504, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 350, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 274, + 365 + ], + "score": 1.0, + "content": "gap to ViT-B can be significantly closed", + "type": "text" + }, + { + "bbox": [ + 275, + 351, + 342, + 362 + ], + "score": 0.77, + "content": "\\mathrm { 5 . 0 \\to 2 . 3 \\ m A P }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 350, + 506, + 365 + ], + "score": 1.0, + "content": ") when fine-tuning with more in-domain", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 361, + 216, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 128, + 374 + ], + "score": 1.0, + "content": "data", + "type": "text" + }, + { + "bbox": [ + 128, + 362, + 210, + 373 + ], + "score": 0.76, + "content": "( \\mathrm { A S } - 2 0 \\mathrm { K } \\mathrm { A S } - 2 \\mathrm { M } )", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 361, + 216, + 374 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 350, + 506, + 374 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 382, + 325, + 523 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 325, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 325, + 394 + ], + "score": 1.0, + "content": "Decoder. Table 1c compares decoder attention types in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 392, + 326, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 326, + 405 + ], + "score": 1.0, + "content": "Audio-MAE. Note that decoders are discarded after pre-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 404, + 325, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 325, + 415 + ], + "score": 1.0, + "content": "training and only the equal-sized ViT-B encoders are", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 415, + 325, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 325, + 425 + ], + "score": 1.0, + "content": "fine-tuned for the end task. Our results show that local", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 425, + 326, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 326, + 437 + ], + "score": 1.0, + "content": "attention with shifted window achieves the best per-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 436, + 326, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 326, + 448 + ], + "score": 1.0, + "content": "formance. Combining local and global attention (i.e.,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 447, + 326, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 326, + 459 + ], + "score": 1.0, + "content": "hybrid attention, Hwin) also improves vanilla global", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 458, + 326, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 326, + 470 + ], + "score": 1.0, + "content": "self-attention. Fig. 5 shows the qualitative reconstruc-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 470, + 326, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 326, + 481 + ], + "score": 1.0, + "content": "tion comparison. In the spectrogram of vowels, the de-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 481, + 325, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 325, + 491 + ], + "score": 1.0, + "content": "coder with local attention reconstructs better harmonics", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 491, + 325, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 325, + 502 + ], + "score": 1.0, + "content": "and recovers more context in the spectrogram. Similar", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 502, + 325, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 325, + 513 + ], + "score": 1.0, + "content": "phenomena are observed in the frictional sound in the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 513, + 182, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 182, + 524 + ], + "score": 1.0, + "content": "middle consonant.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 381, + 326, + 524 + ] + }, + { + "type": "image", + "bbox": [ + 333, + 395, + 504, + 481 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 333, + 395, + 504, + 481 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 333, + 395, + 504, + 481 + ], + "spans": [ + { + "bbox": [ + 333, + 395, + 504, + 481 + ], + "score": 0.922, + "type": "image", + "image_path": "2eac13304af84b06c5760b4abdf221924cb45af9f00f2bcb95dca8f6af1aa8b1.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 333, + 395, + 504, + 407.2857142857143 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 333, + 407.2857142857143, + 504, + 419.57142857142856 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 333, + 419.57142857142856, + 504, + 431.85714285714283 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 333, + 431.85714285714283, + 504, + 444.1428571428571 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 333, + 444.1428571428571, + 504, + 456.4285714285714 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 333, + 456.4285714285714, + 504, + 468.71428571428567 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 333, + 468.71428571428567, + 504, + 480.99999999999994 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 334, + 487, + 501, + 497 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 332, + 486, + 501, + 499 + ], + "spans": [ + { + "bbox": [ + 332, + 486, + 501, + 499 + ], + "score": 1.0, + "content": "Figure 5: Decoder reconstruction comparison.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + } + ], + "index": 39.0 + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "Table 1d ablates the impact of decoder depth on mAP. A deeper 16-layer decoder achieves better", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "performance against its shallower variants. Note that our decoder uses local window attention by", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 273, + 564 + ], + "score": 1.0, + "content": "default where only a fraction of tokens", + "type": "text" + }, + { + "bbox": [ + 273, + 551, + 293, + 561 + ], + "score": 0.86, + "content": "4 { \\times } 4", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 550, + 372, + 564 + ], + "score": 1.0, + "content": "local windows vs.", + "type": "text" + }, + { + "bbox": [ + 372, + 551, + 397, + 561 + ], + "score": 0.89, + "content": "6 4 \\times 8", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 550, + 506, + 564 + ], + "score": 1.0, + "content": "with global attention) are", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "attended. For global attention we find 8-layer decoders to perform better than 16-layer. Table 1e", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 104, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "compares decoder width (embedding dimension). A 512-dimension decoder achieves a good trade-off", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 584, + 375, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 375, + 595 + ], + "score": 1.0, + "content": "between computation and performance as a wider one is not better.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44.5, + "bbox_fs": [ + 104, + 528, + 506, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 603, + 505, + 669 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 615 + ], + "score": 1.0, + "content": "Pre-training Data and Setup. Table 1f summarizes the impact of pre-training dataset size. Overall", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "score": 1.0, + "content": "the model performance is monotonically increasing when using more data for pre-training. Comparing", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 625, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 208, + 637 + ], + "score": 1.0, + "content": "the performance of using", + "type": "text" + }, + { + "bbox": [ + 208, + 626, + 222, + 636 + ], + "score": 0.85, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 625, + 505, + 637 + ], + "score": 1.0, + "content": "well-annotated AS-20K balanced data to using randomly sampled 20K", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 636, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 648 + ], + "score": 1.0, + "content": "unbalanced data for pre-training, the similar mAPs (39.4 vs 39.6) suggest that the distribution of", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "data classes (balanced vs. unbalanced) is less important for pre-training. Meanwhile, as shown in", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 658, + 478, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 131, + 671 + ], + "score": 1.0, + "content": "Table", + "type": "text" + }, + { + "bbox": [ + 131, + 659, + 142, + 670 + ], + "score": 0.27, + "content": "1 \\mathrm { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 658, + 478, + 671 + ], + "score": 1.0, + "content": ", training for longer is beneficial yet the performance saturates after the 24-th epoch.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50.5, + "bbox_fs": [ + 105, + 604, + 506, + 671 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 691 + ], + "score": 1.0, + "content": "Out-of-domain Pre-training on ImageNet. Initializing audio models from ImageNet pre-trained", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "weights has become popular for audio classification. However, as there are significant discrepancies", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "between image and audio modalities, it is questionable if out-of-domain pre-training benefits audio", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "representation learning. In Table 1h we design 3 scenarios to investigate this for Audio-MAE: (1)", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 55.5, + "bbox_fs": [ + 105, + 678, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 65, + 502, + 274 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 65, + 502, + 274 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 65, + 502, + 274 + ], + "spans": [ + { + "bbox": [ + 108, + 65, + 502, + 274 + ], + "score": 0.972, + "html": "
ModerBackboneP1-DataAS-20KAS-ZMIESC-30SPC-2SPC-1SID
No pre-training
ERANN [58]CNN45.089.2
PANN [59]CNN27.843.183.361.8
In-domain self-supervised pre-training
wav2vec 2.0 [33]TransformerLS96.2*75.2*
HuBERT[35]TransformerLS96.3*81.4*
Conformer [37]ConformerAS=41.188.0=--
SS-AST[18]ViT-BAS+LS31.0188.898.096.064.3
Concurrent MAE-based works
MaskSpec [43]ViT-BAS32.347.189.697.7=
MAE-AST[38]ViT-BAS+LS30.6-90.097.995.863.3
Audio-MAE (global)ViT-BAS36.6±.1146.8±.0693.6±.1198.3±.0697.6±.0694.1±.06
Audio-MAE (local)ViT-BAS37.0±.1147.3±.1194.1±.1098.3±.0696.9±.0094.8± .11
Out-of-domain supervised pre-training
PSLA [30]EffNet [60]IN31.944.4=96.3=
AST[10]DeiT-BIN34.745.988.798.195.541.1
MBT[11]ViT-BIN-21K31.344.31-=
HTS-AT [29]Swin-BIN=47.197.0t98.0
PaSST[28]DeiT-BIN47.196.8†-
", + "type": "table", + "image_path": "e0146f12659732d22bc5a656b8f385365a1efb9fcc9883c795d053f067db9e94.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 65, + 502, + 134.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 134.66666666666669, + 502, + 204.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 204.33333333333337, + 502, + 274.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 278, + 506, + 334 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 278, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 505, + 290 + ], + "score": 1.0, + "content": "Table 2: Comparison with other state-of-the-art models on audio and speech classification", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 288, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 289, + 302 + ], + "score": 1.0, + "content": "tasks. Metrics are mAP for AS and accuracy", + "type": "text" + }, + { + "bbox": [ + 290, + 289, + 306, + 300 + ], + "score": 0.7, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 288, + 506, + 302 + ], + "score": 1.0, + "content": "for ESC/SPC/SID. For pre-training (PT) dataset,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 299, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 506, + 313 + ], + "score": 1.0, + "content": "AS:AudioSet, LS:LibriSpeech, and IN:ImageNet. †: Fine-tuning results with additional supervised", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 311, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 506, + 323 + ], + "score": 1.0, + "content": "training on AS-2M. We gray-out models pre-trained with external non-audio datasets (e.g., ImageNet).", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 322, + 500, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 500, + 334 + ], + "score": 1.0, + "content": "Best single models in AS-2M are compared (no ensembles). *: linear evaluation results from [53].", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "Audio-only pre-training (AS-SSL) from scratch. We consider this the ideal schema for learning audio", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "representations as it is a simple and clean setup that prevents uncontrollable bias transfer from other", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 388, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 505, + 400 + ], + "score": 1.0, + "content": "modalities. (2) Directly using self-supervised ImageNet MAE models (IN-SSL) and its fine-tuned", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 399, + 491, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 491, + 411 + ], + "score": 1.0, + "content": "variant (IN-SL). (3) Audio-MAE self-supervised pre-training on top of these ImageNet weights.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 415, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 415, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 506, + 428 + ], + "score": 1.0, + "content": "The results show that (1) from-scratch audio-only pre-training is the best. For scenarios (2) and (3),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 426, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 506, + 439 + ], + "score": 1.0, + "content": "we observe that ImageNet pre-training alone (2) is not sufficient (especially when the downstream", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 436, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 506, + 451 + ], + "score": 1.0, + "content": "data is smaller, AS-20K), and, in self-supervised pre-training on AudioSet, ImageNet initialization", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "(3) does not help but degrades accuracy. Also in (3), supervised ImageNet pre-training (IN-SL) seems", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "harmful. Consequently, the result suggests that out-of-domain pre-training (i.e., ImageNet) is not", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 470, + 325, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 325, + 482 + ], + "score": 1.0, + "content": "helpful for Audio-MAE, possibly due to domain shift.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 107, + 494, + 289, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 290, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 290, + 507 + ], + "score": 1.0, + "content": "4.5 Comparison with the State-of-the-art", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 505, + 580 + ], + "lines": [ + { + "bbox": [ + 106, + 514, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 505, + 526 + ], + "score": 1.0, + "content": "Table 2 compares Audio-MAE (with 3-run error bars) to prior state-of-the-art. We categorize the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 526, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 537 + ], + "score": 1.0, + "content": "comparison into 3 groups. For fair comparison, our main benchmark is the models in the middle", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "score": 1.0, + "content": "group with self-supervised pre-training on in-domain (audio) datasets (AudioSet and LibriSpeech).", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "For reference we also list other models without pre-training (the top group) and other models with", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "supervised pre-training on out-of-domain ImageNet (the bottom group), where the latter contains", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 569, + 259, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 259, + 581 + ], + "score": 1.0, + "content": "previous best systems on the datasets.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 585, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "Pre-trained on AudioSet, Audio-MAE achieves the best performance across all tasks compared", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 460, + 609 + ], + "score": 1.0, + "content": "to other models with in-domain self-supervised pre-training. On AudioSet-20K, its", + "type": "text" + }, + { + "bbox": [ + 460, + 596, + 505, + 607 + ], + "score": 0.51, + "content": "3 7 . 1 \\ \\mathrm { m A P }", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 606, + 507, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 507, + 621 + ], + "score": 1.0, + "content": "significantly outperforms all other approaches including concurrent works and other models with out-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "of-domain pre-training. On AudioSet-2M and ESC-50, our method also outperforms Conformer [37]", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "and SS-AST [18]. Notably, unlike SS-AST and concurrent MAE-AST [38], which trained with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 638, + 457, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 457, + 654 + ], + "score": 1.0, + "content": "additional 1,000 hours of speech in Librispeech, we use only AudioSet for pre-training.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "In the bottom group of Table 2, Audio-MAE also outperforms previous state-of-the-art models with", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "ImageNet supervised pre-training. Note that the proposed Audio-MAE does not rely on any out-of-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "domain data and labels, nor using knowledge distillation (e.g., DeiT) from additional CNN-based", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 688, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 701 + ], + "score": 1.0, + "content": "models. Also, compared to HTS-AT [29] and PaSST [28], Audio-MAE is trained with audio under", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "16K sampling rate. As experimented in [59], there could be up to 0.4 potential mAP improvement for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 710, + 343, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 343, + 723 + ], + "score": 1.0, + "content": "Audio-MAE if audio with 32K sampling rate are available.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 194, + 60, + 481, + 71 + ], + "lines": [ + { + "bbox": [ + 191, + 56, + 483, + 74 + ], + "spans": [ + { + "bbox": [ + 191, + 56, + 483, + 74 + ], + "score": 1.0, + "content": "Backbone PT-Data AS-20K AS-2M ESC-50 SPC-2 SPC-1 SID", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 111, + 61, + 136, + 70 + ], + "lines": [ + { + "bbox": [ + 109, + 60, + 138, + 72 + ], + "spans": [ + { + "bbox": [ + 109, + 60, + 138, + 72 + ], + "score": 1.0, + "content": "Model", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 65, + 502, + 274 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 65, + 502, + 274 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 65, + 502, + 274 + ], + "spans": [ + { + "bbox": [ + 108, + 65, + 502, + 274 + ], + "score": 0.972, + "html": "
ModerBackboneP1-DataAS-20KAS-ZMIESC-30SPC-2SPC-1SID
No pre-training
ERANN [58]CNN45.089.2
PANN [59]CNN27.843.183.361.8
In-domain self-supervised pre-training
wav2vec 2.0 [33]TransformerLS96.2*75.2*
HuBERT[35]TransformerLS96.3*81.4*
Conformer [37]ConformerAS=41.188.0=--
SS-AST[18]ViT-BAS+LS31.0188.898.096.064.3
Concurrent MAE-based works
MaskSpec [43]ViT-BAS32.347.189.697.7=
MAE-AST[38]ViT-BAS+LS30.6-90.097.995.863.3
Audio-MAE (global)ViT-BAS36.6±.1146.8±.0693.6±.1198.3±.0697.6±.0694.1±.06
Audio-MAE (local)ViT-BAS37.0±.1147.3±.1194.1±.1098.3±.0696.9±.0094.8± .11
Out-of-domain supervised pre-training
PSLA [30]EffNet [60]IN31.944.4=96.3=
AST[10]DeiT-BIN34.745.988.798.195.541.1
MBT[11]ViT-BIN-21K31.344.31-=
HTS-AT [29]Swin-BIN=47.197.0t98.0
PaSST[28]DeiT-BIN47.196.8†-
", + "type": "table", + "image_path": "e0146f12659732d22bc5a656b8f385365a1efb9fcc9883c795d053f067db9e94.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 65, + 502, + 134.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 134.66666666666669, + 502, + 204.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 204.33333333333337, + 502, + 274.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 278, + 506, + 334 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 278, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 505, + 290 + ], + "score": 1.0, + "content": "Table 2: Comparison with other state-of-the-art models on audio and speech classification", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 288, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 289, + 302 + ], + "score": 1.0, + "content": "tasks. Metrics are mAP for AS and accuracy", + "type": "text" + }, + { + "bbox": [ + 290, + 289, + 306, + 300 + ], + "score": 0.7, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 288, + 506, + 302 + ], + "score": 1.0, + "content": "for ESC/SPC/SID. For pre-training (PT) dataset,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 299, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 506, + 313 + ], + "score": 1.0, + "content": "AS:AudioSet, LS:LibriSpeech, and IN:ImageNet. †: Fine-tuning results with additional supervised", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 311, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 506, + 323 + ], + "score": 1.0, + "content": "training on AS-2M. We gray-out models pre-trained with external non-audio datasets (e.g., ImageNet).", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 322, + 500, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 500, + 334 + ], + "score": 1.0, + "content": "Best single models in AS-2M are compared (no ensembles). *: linear evaluation results from [53].", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "Audio-only pre-training (AS-SSL) from scratch. We consider this the ideal schema for learning audio", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "representations as it is a simple and clean setup that prevents uncontrollable bias transfer from other", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 388, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 505, + 400 + ], + "score": 1.0, + "content": "modalities. (2) Directly using self-supervised ImageNet MAE models (IN-SSL) and its fine-tuned", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 399, + 491, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 491, + 411 + ], + "score": 1.0, + "content": "variant (IN-SL). (3) Audio-MAE self-supervised pre-training on top of these ImageNet weights.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 366, + 505, + 411 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 415, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 415, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 506, + 428 + ], + "score": 1.0, + "content": "The results show that (1) from-scratch audio-only pre-training is the best. For scenarios (2) and (3),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 426, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 506, + 439 + ], + "score": 1.0, + "content": "we observe that ImageNet pre-training alone (2) is not sufficient (especially when the downstream", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 436, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 506, + 451 + ], + "score": 1.0, + "content": "data is smaller, AS-20K), and, in self-supervised pre-training on AudioSet, ImageNet initialization", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "(3) does not help but degrades accuracy. Also in (3), supervised ImageNet pre-training (IN-SL) seems", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "harmful. Consequently, the result suggests that out-of-domain pre-training (i.e., ImageNet) is not", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 470, + 325, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 325, + 482 + ], + "score": 1.0, + "content": "helpful for Audio-MAE, possibly due to domain shift.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 415, + 506, + 482 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 494, + 289, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 290, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 290, + 507 + ], + "score": 1.0, + "content": "4.5 Comparison with the State-of-the-art", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 505, + 580 + ], + "lines": [ + { + "bbox": [ + 106, + 514, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 505, + 526 + ], + "score": 1.0, + "content": "Table 2 compares Audio-MAE (with 3-run error bars) to prior state-of-the-art. We categorize the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 526, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 537 + ], + "score": 1.0, + "content": "comparison into 3 groups. For fair comparison, our main benchmark is the models in the middle", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "score": 1.0, + "content": "group with self-supervised pre-training on in-domain (audio) datasets (AudioSet and LibriSpeech).", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "For reference we also list other models without pre-training (the top group) and other models with", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "supervised pre-training on out-of-domain ImageNet (the bottom group), where the latter contains", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 569, + 259, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 259, + 581 + ], + "score": 1.0, + "content": "previous best systems on the datasets.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 514, + 506, + 581 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 585, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "Pre-trained on AudioSet, Audio-MAE achieves the best performance across all tasks compared", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 460, + 609 + ], + "score": 1.0, + "content": "to other models with in-domain self-supervised pre-training. On AudioSet-20K, its", + "type": "text" + }, + { + "bbox": [ + 460, + 596, + 505, + 607 + ], + "score": 0.51, + "content": "3 7 . 1 \\ \\mathrm { m A P }", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 606, + 507, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 507, + 621 + ], + "score": 1.0, + "content": "significantly outperforms all other approaches including concurrent works and other models with out-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "of-domain pre-training. On AudioSet-2M and ESC-50, our method also outperforms Conformer [37]", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "and SS-AST [18]. Notably, unlike SS-AST and concurrent MAE-AST [38], which trained with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 638, + 457, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 457, + 654 + ], + "score": 1.0, + "content": "additional 1,000 hours of speech in Librispeech, we use only AudioSet for pre-training.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 585, + 507, + 654 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "In the bottom group of Table 2, Audio-MAE also outperforms previous state-of-the-art models with", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "ImageNet supervised pre-training. Note that the proposed Audio-MAE does not rely on any out-of-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "domain data and labels, nor using knowledge distillation (e.g., DeiT) from additional CNN-based", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 688, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 701 + ], + "score": 1.0, + "content": "models. Also, compared to HTS-AT [29] and PaSST [28], Audio-MAE is trained with audio under", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "16K sampling rate. As experimented in [59], there could be up to 0.4 potential mAP improvement for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 710, + 343, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 343, + 723 + ], + "score": 1.0, + "content": "Audio-MAE if audio with 32K sampling rate are available.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 655, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 71, + 502, + 339 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 71, + 502, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 71, + 502, + 339 + ], + "spans": [ + { + "bbox": [ + 107, + 71, + 502, + 339 + ], + "score": 0.976, + "type": "image", + "image_path": "6244fd1a7a46c5c39262c96d30a4dfd6f022b65c3eb7f73ba45382c9f667a734.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 71, + 502, + 160.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 160.33333333333331, + 502, + 249.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 249.66666666666663, + 502, + 338.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 345, + 506, + 411 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 344, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 357 + ], + "score": 1.0, + "content": "Figure 6: Spectrogram reconstruction visualizations on the AudioSet eval set. Column-wise", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 356, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 368 + ], + "score": 1.0, + "content": "type: speech, music, event, others. Masking type: (a-d) unstructured (random); (e-h) structured", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 128, + 380 + ], + "score": 1.0, + "content": "(time", + "type": "text" + }, + { + "bbox": [ + 128, + 368, + 136, + 377 + ], + "score": 0.28, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 367, + 249, + 380 + ], + "score": 1.0, + "content": "frequency). Masking Ratio:", + "type": "text" + }, + { + "bbox": [ + 249, + 367, + 269, + 378 + ], + "score": 0.87, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 367, + 506, + 380 + ], + "score": 1.0, + "content": ". In each group, we show the original spectrogram (1, top),", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 378, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 432, + 390 + ], + "score": 1.0, + "content": "masked input (2, middle), and MAE output (3, bottom). The spectrogram size is", + "type": "text" + }, + { + "bbox": [ + 433, + 378, + 477, + 388 + ], + "score": 0.9, + "content": "1 0 2 4 \\times 1 2 8", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 378, + 506, + 390 + ], + "score": 1.0, + "content": "; patch", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 389, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 133, + 400 + ], + "score": 1.0, + "content": "size is", + "type": "text" + }, + { + "bbox": [ + 134, + 389, + 162, + 399 + ], + "score": 0.88, + "content": "1 6 \\times 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 389, + 235, + 400 + ], + "score": 1.0, + "content": ". Each sample has", + "type": "text" + }, + { + "bbox": [ + 235, + 389, + 280, + 399 + ], + "score": 0.89, + "content": "6 4 \\times 8 = 5 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 389, + 353, + 400 + ], + "score": 1.0, + "content": "patches with 154 (", + "type": "text" + }, + { + "bbox": [ + 353, + 389, + 372, + 399 + ], + "score": 0.83, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 389, + 505, + 400 + ], + "score": 1.0, + "content": "masked) patches being visible to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 398, + 500, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 500, + 413 + ], + "score": 1.0, + "content": "Audio-MAE. Please click (1 2 3) for audible .wavs. More audible examples are in Supplementary.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 489 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "For the speech tasks (SPC-1, SPC-2, and SID), Audio-MAE outperforms other models without", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "pre-training (ERANN [58], PANN [59]), supervised (AST) and self-supervised models (SS-AST,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "score": 1.0, + "content": "MAE-AST). We further list other works (marked with *) to include the latest results introduced in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 467, + 504, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 504, + 479 + ], + "score": 1.0, + "content": "the SUPERB [53] benchmark. But note that these results are not strictly comparable since SUPERB", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 478, + 497, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 497, + 491 + ], + "score": 1.0, + "content": "employs linear evaluation where the underlying pre-trained models are not end-to-end fine-tuned.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 494, + 505, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "In summary, with audio-only from-scratch pre-training on AudioSet, our Audio-MAE performs well", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 506, + 305, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 305, + 518 + ], + "score": 1.0, + "content": "for both the audio and speech classification tasks.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 107, + 531, + 390, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 530, + 391, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 391, + 545 + ], + "score": 1.0, + "content": "4.6 Visualization and Audible Examples by Audio-MAE Decoder", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 564 + ], + "score": 1.0, + "content": "For better visualization, we follow MAE [1] to use MSE over non-normalized spectrograms as the self-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 564, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 576 + ], + "score": 1.0, + "content": "supervised objective. We use ViT-L as the Audio-MAE encoder for visualization. Fig. 6 illustrates", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 573, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 588 + ], + "score": 1.0, + "content": "the reconstruction results sampled from the AudioSet-2M eval set. We further reconstruct .wavs using", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 584, + 503, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 503, + 598 + ], + "score": 1.0, + "content": "the Griffin-Lim [61] algorithm, audible under the anonymous links (accessible in respective 1 2 3).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 601, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "As can be seen and heard, for various masking strategies and different sounds, our Audio-MAE", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 613, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 624 + ], + "score": 1.0, + "content": "generates reasonable reconstruction. It works well for noisy event sounds (e.g., the reconstructed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "score": 1.0, + "content": "siren in Fig. 6c-3), as well as speech and music (e.g., the reconstructed singing in Fig. 6b-3). Notably,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 635, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 506, + 647 + ], + "score": 1.0, + "content": "unlike visual contents that are typically scale/translation/position invariant [19], absolute positions", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "and arrangement of spectrogram components are critical for humans to understand sound [62]. For", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 657, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 669 + ], + "score": 1.0, + "content": "example, shifting a pitch will make an audio sounds completely different. Also, phoneme sequences", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "in time are important cues for speech understanding. Consequently, unstructured masking produces", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "better aligned outputs that are closer to the ground-truth (top row in each subfigure) as the model can", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "make better predictions based on nearby spectrogram patches; while structured masking is harder", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "(less accurate or with words missing), especially when masking is performed over the time axis. A", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 711, + 399, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 399, + 723 + ], + "score": 1.0, + "content": "failure example (missing words) is the reconstructed speech in Fig. 6e-3.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 71, + 502, + 339 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 71, + 502, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 71, + 502, + 339 + ], + "spans": [ + { + "bbox": [ + 107, + 71, + 502, + 339 + ], + "score": 0.976, + "type": "image", + "image_path": "6244fd1a7a46c5c39262c96d30a4dfd6f022b65c3eb7f73ba45382c9f667a734.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 71, + 502, + 160.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 160.33333333333331, + 502, + 249.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 249.66666666666663, + 502, + 338.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 345, + 506, + 411 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 344, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 357 + ], + "score": 1.0, + "content": "Figure 6: Spectrogram reconstruction visualizations on the AudioSet eval set. Column-wise", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 356, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 368 + ], + "score": 1.0, + "content": "type: speech, music, event, others. Masking type: (a-d) unstructured (random); (e-h) structured", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 128, + 380 + ], + "score": 1.0, + "content": "(time", + "type": "text" + }, + { + "bbox": [ + 128, + 368, + 136, + 377 + ], + "score": 0.28, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 367, + 249, + 380 + ], + "score": 1.0, + "content": "frequency). Masking Ratio:", + "type": "text" + }, + { + "bbox": [ + 249, + 367, + 269, + 378 + ], + "score": 0.87, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 367, + 506, + 380 + ], + "score": 1.0, + "content": ". In each group, we show the original spectrogram (1, top),", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 378, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 432, + 390 + ], + "score": 1.0, + "content": "masked input (2, middle), and MAE output (3, bottom). The spectrogram size is", + "type": "text" + }, + { + "bbox": [ + 433, + 378, + 477, + 388 + ], + "score": 0.9, + "content": "1 0 2 4 \\times 1 2 8", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 378, + 506, + 390 + ], + "score": 1.0, + "content": "; patch", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 389, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 133, + 400 + ], + "score": 1.0, + "content": "size is", + "type": "text" + }, + { + "bbox": [ + 134, + 389, + 162, + 399 + ], + "score": 0.88, + "content": "1 6 \\times 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 389, + 235, + 400 + ], + "score": 1.0, + "content": ". Each sample has", + "type": "text" + }, + { + "bbox": [ + 235, + 389, + 280, + 399 + ], + "score": 0.89, + "content": "6 4 \\times 8 = 5 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 389, + 353, + 400 + ], + "score": 1.0, + "content": "patches with 154 (", + "type": "text" + }, + { + "bbox": [ + 353, + 389, + 372, + 399 + ], + "score": 0.83, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 389, + 505, + 400 + ], + "score": 1.0, + "content": "masked) patches being visible to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 398, + 500, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 500, + 413 + ], + "score": 1.0, + "content": "Audio-MAE. Please click (1 2 3) for audible .wavs. More audible examples are in Supplementary.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 489 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "For the speech tasks (SPC-1, SPC-2, and SID), Audio-MAE outperforms other models without", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "pre-training (ERANN [58], PANN [59]), supervised (AST) and self-supervised models (SS-AST,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "score": 1.0, + "content": "MAE-AST). We further list other works (marked with *) to include the latest results introduced in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 467, + 504, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 504, + 479 + ], + "score": 1.0, + "content": "the SUPERB [53] benchmark. But note that these results are not strictly comparable since SUPERB", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 478, + 497, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 497, + 491 + ], + "score": 1.0, + "content": "employs linear evaluation where the underlying pre-trained models are not end-to-end fine-tuned.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 434, + 506, + 491 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 494, + 505, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "In summary, with audio-only from-scratch pre-training on AudioSet, our Audio-MAE performs well", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 506, + 305, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 305, + 518 + ], + "score": 1.0, + "content": "for both the audio and speech classification tasks.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 494, + 505, + 518 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 531, + 390, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 530, + 391, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 391, + 545 + ], + "score": 1.0, + "content": "4.6 Visualization and Audible Examples by Audio-MAE Decoder", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 564 + ], + "score": 1.0, + "content": "For better visualization, we follow MAE [1] to use MSE over non-normalized spectrograms as the self-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 564, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 576 + ], + "score": 1.0, + "content": "supervised objective. We use ViT-L as the Audio-MAE encoder for visualization. Fig. 6 illustrates", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 573, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 588 + ], + "score": 1.0, + "content": "the reconstruction results sampled from the AudioSet-2M eval set. We further reconstruct .wavs using", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 584, + 503, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 503, + 598 + ], + "score": 1.0, + "content": "the Griffin-Lim [61] algorithm, audible under the anonymous links (accessible in respective 1 2 3).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 552, + 506, + 598 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 601, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "As can be seen and heard, for various masking strategies and different sounds, our Audio-MAE", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 613, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 624 + ], + "score": 1.0, + "content": "generates reasonable reconstruction. It works well for noisy event sounds (e.g., the reconstructed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "score": 1.0, + "content": "siren in Fig. 6c-3), as well as speech and music (e.g., the reconstructed singing in Fig. 6b-3). Notably,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 635, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 506, + 647 + ], + "score": 1.0, + "content": "unlike visual contents that are typically scale/translation/position invariant [19], absolute positions", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "and arrangement of spectrogram components are critical for humans to understand sound [62]. For", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 657, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 669 + ], + "score": 1.0, + "content": "example, shifting a pitch will make an audio sounds completely different. Also, phoneme sequences", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "in time are important cues for speech understanding. Consequently, unstructured masking produces", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "better aligned outputs that are closer to the ground-truth (top row in each subfigure) as the model can", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "make better predictions based on nearby spectrogram patches; while structured masking is harder", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "(less accurate or with words missing), especially when masking is performed over the time axis. A", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 711, + 399, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 399, + 723 + ], + "score": 1.0, + "content": "failure example (missing words) is the reconstructed speech in Fig. 6e-3.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 601, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 183, + 84 + ], + "lines": [ + { + "bbox": [ + 104, + 69, + 185, + 87 + ], + "spans": [ + { + "bbox": [ + 104, + 69, + 185, + 87 + ], + "score": 1.0, + "content": "5 Conclusion", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 95, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 106, + 96, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 106, + 96, + 506, + 108 + ], + "score": 1.0, + "content": "We have explored a simple extension of MAE [1] to audio data. Our Audio-MAE learns to reconstruct", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "masked spectrogram patches from audio recordings and achieves state-of-the-art performance on", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 117, + 506, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 506, + 130 + ], + "score": 1.0, + "content": "six audio and speech classification tasks. We have drawn four interesting observations: First, a", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 128, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 506, + 141 + ], + "score": 1.0, + "content": "simple MAE approach works surprisingly well for audio spectrograms. Second, we find that it is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 139, + 506, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 506, + 152 + ], + "score": 1.0, + "content": "possible to learn stronger representations with local self-attention in the decoder. Third, we show", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 150, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 505, + 163 + ], + "score": 1.0, + "content": "that masking can be applied to both pre-training and fine-tuning, improving accuracy and reducing", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 161, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 505, + 173 + ], + "score": 1.0, + "content": "training computation. The optimal strategy depends on the nature of the data (audio, image, etc.) and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 186 + ], + "score": 1.0, + "content": "the learning type (self-/supervised). Fourth, the best performance can be achieved by pre-training", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 196 + ], + "score": 1.0, + "content": "and fine-tuning under the same modality, without reliance on cross-modality transfer learning. In", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "future work, we aim to explore multimodal self-supervised learning with a joint audio-visual MAE", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 205, + 394, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 394, + 217 + ], + "score": 1.0, + "content": "approach as these domains share natural correspondences in video data.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 224, + 504, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 237 + ], + "score": 1.0, + "content": "Acknowledgements. We thank Kaiming He and Luke Zettlemoyer for their feedback and discussions.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 107, + 251, + 163, + 264 + ], + "lines": [ + { + "bbox": [ + 106, + 250, + 165, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 165, + 266 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 109, + 265, + 506, + 726 + ], + "lines": [ + { + "bbox": [ + 111, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 111, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "[1] K. He, X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick, “Masked autoencoders are scalable", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 282, + 357, + 295 + ], + "spans": [ + { + "bbox": [ + 127, + 282, + 357, + 295 + ], + "score": 1.0, + "content": "vision learners,” arXiv preprint arXiv:2111.06377, 2021.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 111, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 111, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "[2] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 307, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 126, + 307, + 506, + 322 + ], + "score": 1.0, + "content": "I. Polosukhin, “Attention is all you need,” in Proceedings of the 31st International Conference", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 127, + 319, + 507, + 332 + ], + "spans": [ + { + "bbox": [ + 127, + 319, + 507, + 332 + ], + "score": 1.0, + "content": "on Neural Information Processing Systems, ser. NIPS’17. USA: Curran Associates Inc., 2017,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 330, + 193, + 342 + ], + "spans": [ + { + "bbox": [ + 126, + 330, + 193, + 342 + ], + "score": 1.0, + "content": "pp. 6000–6010.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 109, + 343, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 109, + 343, + 506, + 357 + ], + "score": 1.0, + "content": "[3] J. Devlin, M. Chang, K. Lee, and K. Toutanova, “BERT: pre-training of deep bidirectional", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 355, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 127, + 355, + 505, + 368 + ], + "score": 1.0, + "content": "transformers for language understanding,” in Proceedings of the 2019 Conference of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 126, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "North American Chapter of the Association for Computational Linguistics: Human Language", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 377, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 127, + 377, + 506, + 390 + ], + "score": 1.0, + "content": "Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 388, + 459, + 401 + ], + "spans": [ + { + "bbox": [ + 127, + 388, + 459, + 401 + ], + "score": 1.0, + "content": "Short Papers). Association for Computational Linguistics, 2019, pp. 4171–4186.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 110, + 402, + 507, + 416 + ], + "spans": [ + { + "bbox": [ + 110, + 402, + 507, + 416 + ], + "score": 1.0, + "content": "[4] T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 125, + 412, + 507, + 427 + ], + "spans": [ + { + "bbox": [ + 125, + 412, + 507, + 427 + ], + "score": 1.0, + "content": "P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 423, + 507, + 437 + ], + "spans": [ + { + "bbox": [ + 126, + 423, + 507, + 437 + ], + "score": 1.0, + "content": "R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 435, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 127, + 435, + 506, + 447 + ], + "score": 1.0, + "content": "S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 125, + 444, + 507, + 461 + ], + "spans": [ + { + "bbox": [ + 125, + 444, + 507, + 461 + ], + "score": 1.0, + "content": "“Language models are few-shot learners,” in Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 127, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 127, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 127, + 468, + 301, + 480 + ], + "spans": [ + { + "bbox": [ + 127, + 468, + 301, + 480 + ], + "score": 1.0, + "content": "2020, December 6-12, 2020, virtual, 2020.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 111, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 111, + 482, + 506, + 495 + ], + "score": 1.0, + "content": "[5] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 127, + 492, + 507, + 506 + ], + "spans": [ + { + "bbox": [ + 127, + 492, + 507, + 506 + ], + "score": 1.0, + "content": "and V. Stoyanov, “Roberta: A robustly optimized BERT pretraining approach,” CoRR, vol.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 504, + 222, + 516 + ], + "spans": [ + { + "bbox": [ + 127, + 504, + 222, + 516 + ], + "score": 1.0, + "content": "abs/1907.11692, 2019.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 110, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 110, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "[6] K. He, H. Fan, Y. Wu, S. Xie, and R. B. Girshick, “Momentum contrast for unsupervised", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 127, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "visual representation learning,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 540, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 127, + 540, + 506, + 553 + ], + "score": 1.0, + "content": "Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. Computer Vision Foundation /", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 127, + 551, + 244, + 564 + ], + "spans": [ + { + "bbox": [ + 127, + 551, + 244, + 564 + ], + "score": 1.0, + "content": "IEEE, 2020, pp. 9726–9735.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 110, + 565, + 507, + 579 + ], + "spans": [ + { + "bbox": [ + 110, + 565, + 507, + 579 + ], + "score": 1.0, + "content": "[7] X. Chen, S. Xie, and K. He, “An empirical study of training self-supervised vision transformers,”", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 126, + 575, + 507, + 590 + ], + "spans": [ + { + "bbox": [ + 126, + 575, + 507, + 590 + ], + "score": 1.0, + "content": "in 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 128, + 587, + 376, + 600 + ], + "spans": [ + { + "bbox": [ + 128, + 587, + 376, + 600 + ], + "score": 1.0, + "content": "Canada, October 10-17, 2021. IEEE, 2021, pp. 9620–9629.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 111, + 601, + 507, + 615 + ], + "spans": [ + { + "bbox": [ + 111, + 601, + 507, + 615 + ], + "score": 1.0, + "content": "[8] R. Paulus, C. Xiong, and R. Socher, “A deep reinforced model for abstractive summarization,”", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 613, + 266, + 626 + ], + "spans": [ + { + "bbox": [ + 127, + 613, + 266, + 626 + ], + "score": 1.0, + "content": "arXiv, vol. abs/1705.04304, 2017.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 111, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 111, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "[9] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 127, + 639, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 127, + 639, + 506, + 651 + ], + "score": 1.0, + "content": "M. Minderer, G. Heigold, S. Gelly et al., “An image is worth 16x16 words: Transformers for", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 126, + 649, + 403, + 663 + ], + "spans": [ + { + "bbox": [ + 126, + 649, + 403, + 663 + ], + "score": 1.0, + "content": "image recognition at scale,” arXiv preprint arXiv:2010.11929, 2020.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 662, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 104, + 662, + 506, + 678 + ], + "score": 1.0, + "content": "[10] Y. Gong, Y. Chung, and J. R. Glass, “AST: audio spectrogram transformer,” in Interspeech", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 126, + 673, + 507, + 689 + ], + "spans": [ + { + "bbox": [ + 126, + 673, + 507, + 689 + ], + "score": 1.0, + "content": "2021, 22nd Annual Conference of the International Speech Communication Association, Brno,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 127, + 685, + 406, + 699 + ], + "spans": [ + { + "bbox": [ + 127, + 685, + 406, + 699 + ], + "score": 1.0, + "content": "Czechia, 30 August - 3 September 2021. ISCA, 2021, pp. 571–575.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "[11] A. Nagrani, S. Yang, A. Arnab, A. Jansen, C. Schmid, and C. Sun, “Attention bottlenecks", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 126, + 709, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 126, + 709, + 506, + 724 + ], + "score": 1.0, + "content": "for multimodal fusion,” in Advances in Neural Information Processing Systems 34: Annual", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 32.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 183, + 84 + ], + "lines": [ + { + "bbox": [ + 104, + 69, + 185, + 87 + ], + "spans": [ + { + "bbox": [ + 104, + 69, + 185, + 87 + ], + "score": 1.0, + "content": "5 Conclusion", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 95, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 106, + 96, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 106, + 96, + 506, + 108 + ], + "score": 1.0, + "content": "We have explored a simple extension of MAE [1] to audio data. Our Audio-MAE learns to reconstruct", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "masked spectrogram patches from audio recordings and achieves state-of-the-art performance on", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 117, + 506, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 506, + 130 + ], + "score": 1.0, + "content": "six audio and speech classification tasks. We have drawn four interesting observations: First, a", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 128, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 506, + 141 + ], + "score": 1.0, + "content": "simple MAE approach works surprisingly well for audio spectrograms. Second, we find that it is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 139, + 506, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 506, + 152 + ], + "score": 1.0, + "content": "possible to learn stronger representations with local self-attention in the decoder. Third, we show", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 150, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 505, + 163 + ], + "score": 1.0, + "content": "that masking can be applied to both pre-training and fine-tuning, improving accuracy and reducing", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 161, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 505, + 173 + ], + "score": 1.0, + "content": "training computation. The optimal strategy depends on the nature of the data (audio, image, etc.) and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 186 + ], + "score": 1.0, + "content": "the learning type (self-/supervised). Fourth, the best performance can be achieved by pre-training", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 196 + ], + "score": 1.0, + "content": "and fine-tuning under the same modality, without reliance on cross-modality transfer learning. In", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "future work, we aim to explore multimodal self-supervised learning with a joint audio-visual MAE", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 205, + 394, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 394, + 217 + ], + "score": 1.0, + "content": "approach as these domains share natural correspondences in video data.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 96, + 506, + 217 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 224, + 504, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 237 + ], + "score": 1.0, + "content": "Acknowledgements. We thank Kaiming He and Luke Zettlemoyer for their feedback and discussions.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12, + "bbox_fs": [ + 106, + 223, + 505, + 237 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 251, + 163, + 264 + ], + "lines": [ + { + "bbox": [ + 106, + 250, + 165, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 165, + 266 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "list", + "bbox": [ + 109, + 265, + 506, + 726 + ], + "lines": [ + { + "bbox": [ + 111, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 111, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "[1] K. He, X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick, “Masked autoencoders are scalable", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 282, + 357, + 295 + ], + "spans": [ + { + "bbox": [ + 127, + 282, + 357, + 295 + ], + "score": 1.0, + "content": "vision learners,” arXiv preprint arXiv:2111.06377, 2021.", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 111, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "[2] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 307, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 126, + 307, + 506, + 322 + ], + "score": 1.0, + "content": "I. Polosukhin, “Attention is all you need,” in Proceedings of the 31st International Conference", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 127, + 319, + 507, + 332 + ], + "spans": [ + { + "bbox": [ + 127, + 319, + 507, + 332 + ], + "score": 1.0, + "content": "on Neural Information Processing Systems, ser. NIPS’17. USA: Curran Associates Inc., 2017,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 330, + 193, + 342 + ], + "spans": [ + { + "bbox": [ + 126, + 330, + 193, + 342 + ], + "score": 1.0, + "content": "pp. 6000–6010.", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 343, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 109, + 343, + 506, + 357 + ], + "score": 1.0, + "content": "[3] J. Devlin, M. Chang, K. Lee, and K. Toutanova, “BERT: pre-training of deep bidirectional", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 355, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 127, + 355, + 505, + 368 + ], + "score": 1.0, + "content": "transformers for language understanding,” in Proceedings of the 2019 Conference of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 126, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "North American Chapter of the Association for Computational Linguistics: Human Language", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 377, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 127, + 377, + 506, + 390 + ], + "score": 1.0, + "content": "Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 388, + 459, + 401 + ], + "spans": [ + { + "bbox": [ + 127, + 388, + 459, + 401 + ], + "score": 1.0, + "content": "Short Papers). Association for Computational Linguistics, 2019, pp. 4171–4186.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 402, + 507, + 416 + ], + "spans": [ + { + "bbox": [ + 110, + 402, + 507, + 416 + ], + "score": 1.0, + "content": "[4] T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan,", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 412, + 507, + 427 + ], + "spans": [ + { + "bbox": [ + 125, + 412, + 507, + 427 + ], + "score": 1.0, + "content": "P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 423, + 507, + 437 + ], + "spans": [ + { + "bbox": [ + 126, + 423, + 507, + 437 + ], + "score": 1.0, + "content": "R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 435, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 127, + 435, + 506, + 447 + ], + "score": 1.0, + "content": "S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 125, + 444, + 507, + 461 + ], + "spans": [ + { + "bbox": [ + 125, + 444, + 507, + 461 + ], + "score": 1.0, + "content": "“Language models are few-shot learners,” in Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 127, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 127, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 127, + 468, + 301, + 480 + ], + "spans": [ + { + "bbox": [ + 127, + 468, + 301, + 480 + ], + "score": 1.0, + "content": "2020, December 6-12, 2020, virtual, 2020.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 111, + 482, + 506, + 495 + ], + "score": 1.0, + "content": "[5] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 127, + 492, + 507, + 506 + ], + "spans": [ + { + "bbox": [ + 127, + 492, + 507, + 506 + ], + "score": 1.0, + "content": "and V. Stoyanov, “Roberta: A robustly optimized BERT pretraining approach,” CoRR, vol.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 504, + 222, + 516 + ], + "spans": [ + { + "bbox": [ + 127, + 504, + 222, + 516 + ], + "score": 1.0, + "content": "abs/1907.11692, 2019.", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 110, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "[6] K. He, H. Fan, Y. Wu, S. Xie, and R. B. Girshick, “Momentum contrast for unsupervised", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 127, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "visual representation learning,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 540, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 127, + 540, + 506, + 553 + ], + "score": 1.0, + "content": "Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. Computer Vision Foundation /", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 127, + 551, + 244, + 564 + ], + "spans": [ + { + "bbox": [ + 127, + 551, + 244, + 564 + ], + "score": 1.0, + "content": "IEEE, 2020, pp. 9726–9735.", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 565, + 507, + 579 + ], + "spans": [ + { + "bbox": [ + 110, + 565, + 507, + 579 + ], + "score": 1.0, + "content": "[7] X. Chen, S. Xie, and K. He, “An empirical study of training self-supervised vision transformers,”", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 575, + 507, + 590 + ], + "spans": [ + { + "bbox": [ + 126, + 575, + 507, + 590 + ], + "score": 1.0, + "content": "in 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 128, + 587, + 376, + 600 + ], + "spans": [ + { + "bbox": [ + 128, + 587, + 376, + 600 + ], + "score": 1.0, + "content": "Canada, October 10-17, 2021. IEEE, 2021, pp. 9620–9629.", + "type": "text" + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 601, + 507, + 615 + ], + "spans": [ + { + "bbox": [ + 111, + 601, + 507, + 615 + ], + "score": 1.0, + "content": "[8] R. Paulus, C. Xiong, and R. Socher, “A deep reinforced model for abstractive summarization,”", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 613, + 266, + 626 + ], + "spans": [ + { + "bbox": [ + 127, + 613, + 266, + 626 + ], + "score": 1.0, + "content": "arXiv, vol. abs/1705.04304, 2017.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 111, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "[9] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 127, + 639, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 127, + 639, + 506, + 651 + ], + "score": 1.0, + "content": "M. Minderer, G. Heigold, S. Gelly et al., “An image is worth 16x16 words: Transformers for", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 126, + 649, + 403, + 663 + ], + "spans": [ + { + "bbox": [ + 126, + 649, + 403, + 663 + ], + "score": 1.0, + "content": "image recognition at scale,” arXiv preprint arXiv:2010.11929, 2020.", + "type": "text" + } + ], + "index": 46, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 662, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 104, + 662, + 506, + 678 + ], + "score": 1.0, + "content": "[10] Y. Gong, Y. Chung, and J. R. Glass, “AST: audio spectrogram transformer,” in Interspeech", + "type": "text" + } + ], + "index": 47, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 673, + 507, + 689 + ], + "spans": [ + { + "bbox": [ + 126, + 673, + 507, + 689 + ], + "score": 1.0, + "content": "2021, 22nd Annual Conference of the International Speech Communication Association, Brno,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 127, + 685, + 406, + 699 + ], + "spans": [ + { + "bbox": [ + 127, + 685, + 406, + 699 + ], + "score": 1.0, + "content": "Czechia, 30 August - 3 September 2021. ISCA, 2021, pp. 571–575.", + "type": "text" + } + ], + "index": 49, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "[11] A. Nagrani, S. Yang, A. Arnab, A. Jansen, C. Schmid, and C. Sun, “Attention bottlenecks", + "type": "text" + } + ], + "index": 50, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 709, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 126, + 709, + 506, + 724 + ], + "score": 1.0, + "content": "for multimodal fusion,” in Advances in Neural Information Processing Systems 34: Annual", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 127, + 72, + 507, + 87 + ], + "spans": [ + { + "bbox": [ + 127, + 72, + 507, + 87 + ], + "score": 1.0, + "content": "Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14,", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 82, + 289, + 99 + ], + "spans": [ + { + "bbox": [ + 126, + 82, + 289, + 99 + ], + "score": 1.0, + "content": "2021, virtual, 2021, pp. 14 200–14 213.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 99, + 506, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 506, + 113 + ], + "score": 1.0, + "content": "[12] J. F. Gemmeke, D. P. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 126, + 110, + 506, + 124 + ], + "score": 1.0, + "content": "M. Ritter, “Audio set: An ontology and human-labeled dataset for audio events,” in 2017 IEEE", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 121, + 507, + 135 + ], + "spans": [ + { + "bbox": [ + 126, + 121, + 507, + 135 + ], + "score": 1.0, + "content": "International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2017,", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 132, + 184, + 145 + ], + "spans": [ + { + "bbox": [ + 125, + 132, + 184, + 145 + ], + "score": 1.0, + "content": "pp. 776–780.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 145, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 104, + 145, + 505, + 162 + ], + "score": 1.0, + "content": "[13] K. J. Piczak, “ESC: Dataset for Environmental Sound Classification,” in Proceedings of the", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 158, + 463, + 172 + ], + "spans": [ + { + "bbox": [ + 126, + 158, + 463, + 172 + ], + "score": 1.0, + "content": "23rd Annual ACM Conference on Multimedia. ACM Press, 2015, pp. 1015–1018.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "score": 1.0, + "content": "[14] H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jégou, “Training data-", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 186, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 127, + 186, + 505, + 198 + ], + "score": 1.0, + "content": "efficient image transformers & distillation through attention,” in International Conference on", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 196, + 349, + 209 + ], + "spans": [ + { + "bbox": [ + 125, + 196, + 349, + 209 + ], + "score": 1.0, + "content": "Machine Learning. PMLR, 2021, pp. 10 347–10 357.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "[15] J. Carreira and A. Zisserman, “Quo vadis, action recognition? A new model and the kinetics", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 224, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 127, + 224, + 506, + 236 + ], + "score": 1.0, + "content": "dataset,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017,", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 234, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 127, + 234, + 479, + 248 + ], + "score": 1.0, + "content": "Honolulu, HI, USA, July 21-26, 2017. IEEE Computer Society, 2017, pp. 4724–4733.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 249, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 264 + ], + "score": 1.0, + "content": "[16] C. Feichtenhofer, A. Pinz, and R. P. Wildes, “Spatiotemporal residual networks for video action", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 260, + 244, + 273 + ], + "spans": [ + { + "bbox": [ + 127, + 260, + 244, + 273 + ], + "score": 1.0, + "content": "recognition,” in NIPS, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "score": 1.0, + "content": "[17] H. Bao, L. Dong, and F. Wei, “Beit: BERT pre-training of image transformers,” CoRR, vol.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 287, + 222, + 299 + ], + "spans": [ + { + "bbox": [ + 127, + 287, + 222, + 299 + ], + "score": 1.0, + "content": "abs/2106.08254, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 302, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 317 + ], + "score": 1.0, + "content": "[18] Y. Gong, C.-I. Lai, Y.-A. Chung, and J. R. Glass, “Ssast: Self-supervised audio spectrogram", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 313, + 324, + 327 + ], + "spans": [ + { + "bbox": [ + 127, + 313, + 324, + 327 + ], + "score": 1.0, + "content": "transformer,” ArXiv, vol. abs/2110.09784, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 328, + 507, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 507, + 343 + ], + "score": 1.0, + "content": "[19] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer:", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 341, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 127, + 341, + 505, + 353 + ], + "score": 1.0, + "content": "Hierarchical vision transformer using shifted windows,” in 2021 IEEE/CVF International", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 351, + 507, + 365 + ], + "spans": [ + { + "bbox": [ + 127, + 351, + 507, + 365 + ], + "score": 1.0, + "content": "Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 361, + 252, + 376 + ], + "spans": [ + { + "bbox": [ + 127, + 361, + 252, + 376 + ], + "score": 1.0, + "content": "IEEE, 2021, pp. 9992–10 002.", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 376, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 506, + 391 + ], + "score": 1.0, + "content": "[20] C. Wei, H. Fan, S. Xie, C. Wu, A. L. Yuille, and C. Feichtenhofer, “Masked feature prediction", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 388, + 427, + 402 + ], + "spans": [ + { + "bbox": [ + 127, + 388, + 427, + 402 + ], + "score": 1.0, + "content": "for self-supervised visual pre-training,” CoRR, vol. abs/2112.09133, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 404, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 506, + 418 + ], + "score": 1.0, + "content": "[21] P. Vincent, H. Larochelle, Y. Bengio, and P. Manzagol, “Extracting and composing robust", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 415, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 127, + 415, + 505, + 427 + ], + "score": 1.0, + "content": "features with denoising autoencoders,” in Machine Learning, Proceedings of the Twenty-Fifth", + "type": "text", + "cross_page": true + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 426, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 127, + 426, + 506, + 440 + ], + "score": 1.0, + "content": "International Conference (ICML 2008), Helsinki, Finland, June 5-9, 2008, ser. ACM", + "type": "text", + "cross_page": true + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 438, + 468, + 450 + ], + "spans": [ + { + "bbox": [ + 127, + 438, + 468, + 450 + ], + "score": 1.0, + "content": "International Conference Proceeding Series, vol. 307. ACM, 2008, pp. 1096–1103.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 451, + 507, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 507, + 468 + ], + "score": 1.0, + "content": "[22] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P. Manzagol, “Stacked denoising", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 461, + 507, + 479 + ], + "spans": [ + { + "bbox": [ + 124, + 461, + 507, + 479 + ], + "score": 1.0, + "content": "autoencoders: Learning useful representations in a deep network with a local denoising", + "type": "text", + "cross_page": true + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 474, + 380, + 488 + ], + "spans": [ + { + "bbox": [ + 126, + 474, + 380, + 488 + ], + "score": 1.0, + "content": "criterion,” J. Mach. Learn. Res., vol. 11, pp. 3371–3408, 2010.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 490, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 505 + ], + "score": 1.0, + "content": "[23] D. Pathak, P. Krähenbühl, J. Donahue, T. Darrell, and A. A. Efros, “Context encoders: Feature", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 501, + 507, + 515 + ], + "spans": [ + { + "bbox": [ + 127, + 501, + 507, + 515 + ], + "score": 1.0, + "content": "learning by inpainting,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition,", + "type": "text", + "cross_page": true + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 509, + 508, + 528 + ], + "spans": [ + { + "bbox": [ + 126, + 509, + 508, + 528 + ], + "score": 1.0, + "content": "CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016. IEEE Computer Society, 2016, pp.", + "type": "text", + "cross_page": true + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 523, + 181, + 536 + ], + "spans": [ + { + "bbox": [ + 127, + 523, + 181, + 536 + ], + "score": 1.0, + "content": "2536–2544.", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 539, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 505, + 551 + ], + "score": 1.0, + "content": "[24] M. Chen, A. Radford, R. Child, J. Wu, H. Jun, D. Luan, and I. Sutskever, “Generative", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 548, + 507, + 564 + ], + "spans": [ + { + "bbox": [ + 125, + 548, + 507, + 564 + ], + "score": 1.0, + "content": "pretraining from pixels,” in Proceedings of the 37th International Conference on Machine", + "type": "text", + "cross_page": true + } + ], + "index": 38 + }, + { + "bbox": [ + 125, + 558, + 507, + 576 + ], + "spans": [ + { + "bbox": [ + 125, + 558, + 507, + 576 + ], + "score": 1.0, + "content": "Learning, ICML 2020, 13-18 July 2020, Virtual Event, ser. Proceedings of Machine Learning", + "type": "text", + "cross_page": true + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 571, + 337, + 585 + ], + "spans": [ + { + "bbox": [ + 127, + 571, + 337, + 585 + ], + "score": 1.0, + "content": "Research, vol. 119. PMLR, 2020, pp. 1691–1703.", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 588, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 506, + 600 + ], + "score": 1.0, + "content": "[25] A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. Voss, A. Radford, M. Chen, and I. Sutskever,", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 127, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "“Zero-shot text-to-image generation,” in Proceedings of the 38th International Conference on", + "type": "text", + "cross_page": true + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 127, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, ser. Proceedings of Machine", + "type": "text", + "cross_page": true + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 619, + 376, + 634 + ], + "spans": [ + { + "bbox": [ + 127, + 619, + 376, + 634 + ], + "score": 1.0, + "content": "Learning Research, vol. 139. PMLR, 2021, pp. 8821–8831.", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "[26] N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in 2005 IEEE", + "type": "text", + "cross_page": true + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 648, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 127, + 648, + 505, + 659 + ], + "score": 1.0, + "content": "Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2005),", + "type": "text", + "cross_page": true + } + ], + "index": 46 + }, + { + "bbox": [ + 126, + 656, + 475, + 672 + ], + "spans": [ + { + "bbox": [ + 126, + 656, + 475, + 672 + ], + "score": 1.0, + "content": "20-26 June 2005, San Diego, CA, USA. IEEE Computer Society, 2005, pp. 886–893.", + "type": "text", + "cross_page": true + } + ], + "index": 47, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 673, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 506, + 687 + ], + "score": 1.0, + "content": "[27] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” 2016", + "type": "text", + "cross_page": true + } + ], + "index": 48, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 684, + 499, + 698 + ], + "spans": [ + { + "bbox": [ + 127, + 684, + 499, + 698 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 698, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 714 + ], + "score": 1.0, + "content": "[28] K. Koutini, J. Schlüter, H. Eghbal-zadeh, and G. Widmer, “Efficient training of audio", + "type": "text", + "cross_page": true + } + ], + "index": 50, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 709, + 385, + 725 + ], + "spans": [ + { + "bbox": [ + 125, + 709, + 385, + 725 + ], + "score": 1.0, + "content": "transformers with patchout,” CoRR, vol. abs/2110.05069, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 51, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "[29] K. Chen, X. Du, B. Zhu, Z. Ma, T. Berg-Kirkpatrick, and S. Dubnov, “Hts-at: A hierarchi-", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 83, + 506, + 98 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 506, + 98 + ], + "score": 1.0, + "content": "cal token-semantic audio transformer for sound classification and detection,” arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 94, + 231, + 107 + ], + "spans": [ + { + "bbox": [ + 126, + 94, + 231, + 107 + ], + "score": 1.0, + "content": "arXiv:2202.00874, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 110, + 507, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 507, + 124 + ], + "score": 1.0, + "content": "[30] Y. Gong, Y. Chung, and J. Glass, “Psla: Improving audio tagging with pretraining, sampling, la-", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 118, + 507, + 136 + ], + "spans": [ + { + "bbox": [ + 125, + 118, + 507, + 136 + ], + "score": 1.0, + "content": "beling, and aggregation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing,", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 129, + 156, + 146 + ], + "spans": [ + { + "bbox": [ + 126, + 129, + 156, + 146 + ], + "score": 1.0, + "content": "2021.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "score": 1.0, + "content": "[31] Y. Gong, S. Khurana, A. Rouditchenko, and J. Glass, “Cmkd: Cnn/transformer-based", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 157, + 394, + 171 + ], + "spans": [ + { + "bbox": [ + 126, + 157, + 394, + 171 + ], + "score": 1.0, + "content": "cross-model knowledge distillation for audio classification,” 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 171, + 507, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 507, + 189 + ], + "score": 1.0, + "content": "[32] S. Schneider, A. Baevski, R. Collobert, and M. Auli, “wav2vec: Unsupervised pre-training", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 184, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 127, + 184, + 506, + 198 + ], + "score": 1.0, + "content": "for speech recognition,” in Interspeech 2019, 20th Annual Conference of the International", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 193, + 507, + 211 + ], + "spans": [ + { + "bbox": [ + 125, + 193, + 507, + 211 + ], + "score": 1.0, + "content": "Speech Communication Association, Graz, Austria, 15-19 September 2019. ISCA, 2019, pp.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 206, + 179, + 220 + ], + "spans": [ + { + "bbox": [ + 127, + 206, + 179, + 220 + ], + "score": 1.0, + "content": "3465–3469.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 221, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 506, + 235 + ], + "score": 1.0, + "content": "[33] A. Baevski, Y. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 232, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 127, + 232, + 506, + 246 + ], + "score": 1.0, + "content": "self-supervised learning of speech representations,” in Advances in Neural Information", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 127, + 242, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 127, + 242, + 505, + 256 + ], + "score": 1.0, + "content": "Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020,", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 254, + 338, + 268 + ], + "spans": [ + { + "bbox": [ + 126, + 254, + 338, + 268 + ], + "score": 1.0, + "content": "NeurIPS 2020, December 6-12, 2020, virtual, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 268, + 507, + 284 + ], + "spans": [ + { + "bbox": [ + 104, + 268, + 507, + 284 + ], + "score": 1.0, + "content": "[34] A. Baevski, W. Hsu, Q. Xu, A. Babu, J. Gu, and M. Auli, “data2vec: A general framework for", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 280, + 497, + 295 + ], + "spans": [ + { + "bbox": [ + 125, + 280, + 497, + 295 + ], + "score": 1.0, + "content": "self-supervised learning in speech, vision and language,” CoRR, vol. abs/2202.03555, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 296, + 507, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 507, + 310 + ], + "score": 1.0, + "content": "[35] W.-N. Hsu, B. Bolte, Y.-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “Hu-", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 127, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "bert: Self-supervised speech representation learning by masked prediction of hidden units,”", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 125, + 316, + 507, + 333 + ], + "spans": [ + { + "bbox": [ + 125, + 316, + 507, + 333 + ], + "score": 1.0, + "content": "IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 3451–3460,", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 328, + 155, + 343 + ], + "spans": [ + { + "bbox": [ + 126, + 328, + 155, + 343 + ], + "score": 1.0, + "content": "2021.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 344, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 358 + ], + "score": 1.0, + "content": "[36] B. Shi, W. Hsu, and A. Mohamed, “Robust self-supervised audio-visual speech recognition,”", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 355, + 268, + 369 + ], + "spans": [ + { + "bbox": [ + 127, + 355, + 268, + 369 + ], + "score": 1.0, + "content": "CoRR, vol. abs/2201.01763, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 370, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 384 + ], + "score": 1.0, + "content": "[37] S. Srivastava, Y. Wang, A. Tjandra, A. Kumar, C. Liu, K. Singh, and Y. Saraf, “Conformer-based", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 381, + 504, + 396 + ], + "spans": [ + { + "bbox": [ + 127, + 381, + 504, + 396 + ], + "score": 1.0, + "content": "self-supervised learning for non-speech audio tasks,” arXiv preprint arXiv:2110.07313, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 395, + 507, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 507, + 412 + ], + "score": 1.0, + "content": "[38] A. Baade, P. Peng, and D. Harwath, “Mae-ast: Masked autoencoding audio spectrogram", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 408, + 345, + 421 + ], + "spans": [ + { + "bbox": [ + 126, + 408, + 345, + 421 + ], + "score": 1.0, + "content": "transformer,” arXiv preprint arXiv:2203.16691, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "score": 1.0, + "content": "[39] A. van den Oord, Y. Li, and O. Vinyals, “Representation learning with contrastive predictive", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 433, + 304, + 448 + ], + "spans": [ + { + "bbox": [ + 126, + 433, + 304, + 448 + ], + "score": 1.0, + "content": "coding,” CoRR, vol. abs/1807.03748, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 449, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 464 + ], + "score": 1.0, + "content": "[40] R. Arandjelovic and A. Zisserman, “Objects that sound,” in Computer Vision - ECCV 2018 -", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 460, + 507, + 475 + ], + "spans": [ + { + "bbox": [ + 127, + 460, + 507, + 475 + ], + "score": 1.0, + "content": "15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part I, ser.", + "type": "text", + "cross_page": true + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 472, + 449, + 486 + ], + "spans": [ + { + "bbox": [ + 127, + 472, + 449, + 486 + ], + "score": 1.0, + "content": "Lecture Notes in Computer Science, vol. 11205. Springer, 2018, pp. 451–466.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 487, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 507, + 501 + ], + "score": 1.0, + "content": "[41] M. Patrick, P. Huang, I. Misra, F. Metze, A. Vedaldi, Y. M. Asano, and J. F. Henriques,", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 497, + 506, + 512 + ], + "spans": [ + { + "bbox": [ + 126, + 497, + 506, + 512 + ], + "score": 1.0, + "content": "“Space-time crop & attend: Improving cross-modal video representation learning,” in 2021", + "type": "text", + "cross_page": true + } + ], + "index": 34 + }, + { + "bbox": [ + 125, + 507, + 507, + 524 + ], + "spans": [ + { + "bbox": [ + 125, + 507, + 507, + 524 + ], + "score": 1.0, + "content": "IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada,", + "type": "text", + "cross_page": true + } + ], + "index": 35 + }, + { + "bbox": [ + 126, + 519, + 355, + 535 + ], + "spans": [ + { + "bbox": [ + 126, + 519, + 355, + 535 + ], + "score": 1.0, + "content": "October 10-17, 2021. IEEE, 2021, pp. 10 540–10 552.", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 534, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 549 + ], + "score": 1.0, + "content": "[42] A. T. Liu, S. Yang, P. Chi, P. Hsu, and H. Lee, “Mockingjay: Unsupervised speech", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 545, + 507, + 560 + ], + "spans": [ + { + "bbox": [ + 125, + 545, + 507, + 560 + ], + "score": 1.0, + "content": "representation learning with deep bidirectional transformer encoders,” in 2020 IEEE", + "type": "text", + "cross_page": true + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 556, + 507, + 571 + ], + "spans": [ + { + "bbox": [ + 127, + 556, + 507, + 571 + ], + "score": 1.0, + "content": "International Conference on Acoustics, Speech and Signal Processing, ICASSP 2020,", + "type": "text", + "cross_page": true + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 568, + 389, + 582 + ], + "spans": [ + { + "bbox": [ + 127, + 568, + 389, + 582 + ], + "score": 1.0, + "content": "Barcelona, Spain, May 4-8, 2020. IEEE, 2020, pp. 6419–6423.", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 583, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 597 + ], + "score": 1.0, + "content": "[43] D. Chong, H. Wang, P. Zhou, and Q. Zeng, “Masked spectrogram prediction for self-supervised", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 594, + 233, + 608 + ], + "spans": [ + { + "bbox": [ + 127, + 594, + 233, + 608 + ], + "score": 1.0, + "content": "audio pre-training,” 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 609, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 506, + 624 + ], + "score": 1.0, + "content": "[44] D. Niizumi, D. Takeuchi, Y. Ohishi, N. Harada, and K. Kashino, “Masked spectrogram", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 127, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "modeling using masked autoencoders for learning general-purpose audio representation,”", + "type": "text", + "cross_page": true + } + ], + "index": 44 + }, + { + "bbox": [ + 126, + 631, + 231, + 645 + ], + "spans": [ + { + "bbox": [ + 126, + 631, + 231, + 645 + ], + "score": 1.0, + "content": "arXiv:2204.12260, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 45, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 648, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 107, + 648, + 505, + 659 + ], + "score": 1.0, + "content": "[45] Y. Li, C.-Y. Wu, H. Fan, K. Mangalam, B. Xiong, J. Malik, and C. Feichtenhofer, “Mvitv2:", + "type": "text", + "cross_page": true + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 659, + 486, + 671 + ], + "spans": [ + { + "bbox": [ + 127, + 659, + 486, + 671 + ], + "score": 1.0, + "content": "Improved multiscale vision transformers for classification and detection,” in CVPR, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 47, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 673, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 505, + 687 + ], + "score": 1.0, + "content": "[46] A. van den Oord, Y. Li, and O. Vinyals, “Representation learning with contrastive predictive", + "type": "text", + "cross_page": true + } + ], + "index": 48, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 684, + 304, + 698 + ], + "spans": [ + { + "bbox": [ + 127, + 684, + 304, + 698 + ], + "score": 1.0, + "content": "coding,” CoRR, vol. abs/1807.03748, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 49, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 698, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 714 + ], + "score": 1.0, + "content": "[47] A. Baevski, M. Auli, and A. Mohamed, “Effectiveness of self-supervised pre-training for", + "type": "text", + "cross_page": true + } + ], + "index": 50, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 710, + 352, + 724 + ], + "spans": [ + { + "bbox": [ + 126, + 710, + 352, + 724 + ], + "score": 1.0, + "content": "speech recognition,” CoRR, vol. abs/1911.03912, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 51, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 70, + 507, + 89 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 507, + 89 + ], + "score": 1.0, + "content": "[48] D. S. Park, W. Chan, Y. Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V. Le, “Specaug-", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 84, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 127, + 84, + 507, + 97 + ], + "score": 1.0, + "content": "ment: A simple data augmentation method for automatic speech recognition,” ArXiv, vol.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 95, + 221, + 106 + ], + "spans": [ + { + "bbox": [ + 127, + 95, + 221, + 106 + ], + "score": 1.0, + "content": "abs/1904.08779, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "score": 1.0, + "content": "[49] J. B. Li, S. Qu, P. Huang, and F. Metze, “Audiotagging done right: 2nd comparison of deep", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 119, + 502, + 135 + ], + "spans": [ + { + "bbox": [ + 126, + 119, + 502, + 135 + ], + "score": 1.0, + "content": "learning methods for environmental sound classification,” CoRR, vol. abs/2203.13448, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 135, + 507, + 148 + ], + "spans": [ + { + "bbox": [ + 106, + 135, + 507, + 148 + ], + "score": 1.0, + "content": "[50] S. Hershey, S. Chaudhuri, D. P. W. Ellis, J. F. Gemmeke, A. Jansen, C. Moore, M. Plakal,", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 146, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 126, + 146, + 506, + 160 + ], + "score": 1.0, + "content": "D. Platt, R. A. Saurous, B. Seybold, M. Slaney, R. Weiss, and K. Wilson, “Cnn architectures for", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 155, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 126, + 155, + 506, + 172 + ], + "score": 1.0, + "content": "large-scale audio classification,” in International Conference on Acoustics, Speech and Signal", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 167, + 244, + 181 + ], + "spans": [ + { + "bbox": [ + 126, + 167, + 244, + 181 + ], + "score": 1.0, + "content": "Processing (ICASSP), 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 182, + 507, + 198 + ], + "spans": [ + { + "bbox": [ + 104, + 182, + 507, + 198 + ], + "score": 1.0, + "content": "[51] S. Hershey, D. P. Ellis, E. Fonseca, A. Jansen, C. Liu, R. C. Moore, and M. Plakal, “The", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 192, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 126, + 192, + 506, + 207 + ], + "score": 1.0, + "content": "benefit of temporally-strong labels in audio event classification,” in ICASSP 2021-2021 IEEE", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 125, + 203, + 507, + 219 + ], + "spans": [ + { + "bbox": [ + 125, + 203, + 507, + 219 + ], + "score": 1.0, + "content": "International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021,", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 216, + 183, + 228 + ], + "spans": [ + { + "bbox": [ + 126, + 216, + 183, + 228 + ], + "score": 1.0, + "content": "pp. 366–370.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "[52] P. Warden, “Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition,” ArXiv", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 242, + 209, + 254 + ], + "spans": [ + { + "bbox": [ + 126, + 242, + 209, + 254 + ], + "score": 1.0, + "content": "e-prints, Apr. 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 255, + 507, + 271 + ], + "spans": [ + { + "bbox": [ + 104, + 255, + 507, + 271 + ], + "score": 1.0, + "content": "[53] S. wen Yang, P.-H. Chi, Y.-S. Chuang, C.-I. J. Lai, K. Lakhotia, Y. Y. Lin, A. T. Liu, J. Shi,", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 266, + 507, + 282 + ], + "spans": [ + { + "bbox": [ + 124, + 266, + 507, + 282 + ], + "score": 1.0, + "content": "X. Chang, G.-T. Lin, T.-H. Huang, W.-C. Tseng, K. tik Lee, D.-R. Liu, Z. Huang, S. Dong,", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 127, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "S.-W. Li, S. Watanabe, A. Mohamed, and H. yi Lee, “SUPERB: Speech Processing Universal", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 288, + 438, + 302 + ], + "spans": [ + { + "bbox": [ + 126, + 288, + 438, + 302 + ], + "score": 1.0, + "content": "PERformance Benchmark,” in Proc. Interspeech 2021, 2021, pp. 1194–1198.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "[54] A. Nagrani, J. S. Chung, W. Xie, and A. Zisserman, “Voxceleb: Large-scale speaker verification", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 316, + 331, + 327 + ], + "spans": [ + { + "bbox": [ + 127, + 316, + 331, + 327 + ], + "score": 1.0, + "content": "in the wild,” Comput. Speech Lang., vol. 60, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 328, + 507, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 507, + 344 + ], + "score": 1.0, + "content": "[55] D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann,", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 127, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "P. Motlicek, Y. Qian, P. Schwarz et al., “The kaldi speech recognition toolkit,” in IEEE 2011", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 350, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 126, + 350, + 506, + 365 + ], + "score": 1.0, + "content": "workshop on automatic speech recognition and understanding, no. CONF. IEEE Signal", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 362, + 234, + 376 + ], + "spans": [ + { + "bbox": [ + 127, + 362, + 234, + 376 + ], + "score": 1.0, + "content": "Processing Society, 2011.", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 375, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 104, + 375, + 506, + 393 + ], + "score": 1.0, + "content": "[56] S. Yun, D. Han, S. Chun, S. J. Oh, Y. Yoo, and J. Choe, “Cutmix: Regularization strategy to", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 127, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "train strong classifiers with localizable features,” in 2019 IEEE/CVF International Conference", + "type": "text", + "cross_page": true + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 398, + 507, + 413 + ], + "spans": [ + { + "bbox": [ + 126, + 398, + 507, + 413 + ], + "score": 1.0, + "content": "on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 409, + 245, + 424 + ], + "spans": [ + { + "bbox": [ + 127, + 409, + 245, + 424 + ], + "score": 1.0, + "content": "IEEE, 2019, pp. 6022–6031.", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 424, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 439 + ], + "score": 1.0, + "content": "[57] H. Zhang, M. Cissé, Y. N. Dauphin, and D. Lopez-Paz, “mixup: Beyond empirical", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 127, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "risk minimization,” in 6th International Conference on Learning Representations, ICLR", + "type": "text", + "cross_page": true + } + ], + "index": 30 + }, + { + "bbox": [ + 126, + 445, + 507, + 462 + ], + "spans": [ + { + "bbox": [ + 126, + 445, + 507, + 462 + ], + "score": 1.0, + "content": "2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings.", + "type": "text", + "cross_page": true + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 458, + 225, + 471 + ], + "spans": [ + { + "bbox": [ + 127, + 458, + 225, + 471 + ], + "score": 1.0, + "content": "OpenReview.net, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 471, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 104, + 471, + 506, + 486 + ], + "score": 1.0, + "content": "[58] S. Verbitskiy, V. Berikov, and V. Vyshegorodtsev, “Eranns: Efficient residual audio neural", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 484, + 453, + 497 + ], + "spans": [ + { + "bbox": [ + 126, + 484, + 453, + 497 + ], + "score": 1.0, + "content": "networks for audio pattern recognition,” arXiv preprint arXiv:2106.01621, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "[59] Q. Kong, Y. Cao, T. Iqbal, Y. Wang, W. Wang, and M. D. Plumbley, “Panns: Large-scale", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 509, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 126, + 509, + 506, + 523 + ], + "score": 1.0, + "content": "pretrained audio neural networks for audio pattern recognition,” IEEE ACM Trans. Audio", + "type": "text", + "cross_page": true + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 520, + 345, + 534 + ], + "spans": [ + { + "bbox": [ + 127, + 520, + 345, + 534 + ], + "score": 1.0, + "content": "Speech Lang. Process., vol. 28, pp. 2880–2894, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 534, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 104, + 534, + 506, + 549 + ], + "score": 1.0, + "content": "[60] M. Tan and Q. V. Le, “Efficientnet: Rethinking model scaling for convolutional neural", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 127, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "networks,” in Proceedings of the 36th International Conference on Machine Learning, ICML", + "type": "text", + "cross_page": true + } + ], + "index": 39 + }, + { + "bbox": [ + 126, + 555, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 126, + 555, + 506, + 572 + ], + "score": 1.0, + "content": "2019, 9-15 June 2019, Long Beach, California, USA, ser. Proceedings of Machine Learning", + "type": "text", + "cross_page": true + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 567, + 331, + 581 + ], + "spans": [ + { + "bbox": [ + 127, + 567, + 331, + 581 + ], + "score": 1.0, + "content": "Research, vol. 97. PMLR, 2019, pp. 6105–6114.", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "[61] D. Griffin and J. Lim, “Signal estimation from modified short-time fourier transform,” IEEE", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 594, + 502, + 608 + ], + "spans": [ + { + "bbox": [ + 127, + 594, + 502, + 608 + ], + "score": 1.0, + "content": "Transactions on Acoustics, Speech, and Signal Processing, vol. 32, no. 2, pp. 236–243, 1984.", + "type": "text", + "cross_page": true + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "[62] Y. Suzuki and H. Takeshima, “Equal-loudness-level contours for pure tones,” The Journal of the", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 619, + 395, + 633 + ], + "spans": [ + { + "bbox": [ + 126, + 619, + 395, + 633 + ], + "score": 1.0, + "content": "Acoustical Society of America, vol. 116, no. 2, pp. 918–933, 2004.", + "type": "text", + "cross_page": true + } + ], + "index": 45, + "is_list_end_line": true + } + ], + "index": 32.5, + "bbox_fs": [ + 104, + 271, + 507, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 37, + 507, + 727 + ], + "lines": [ + { + "bbox": [ + 127, + 72, + 507, + 87 + ], + "spans": [ + { + "bbox": [ + 127, + 72, + 507, + 87 + ], + "score": 1.0, + "content": "Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 82, + 289, + 99 + ], + "spans": [ + { + "bbox": [ + 126, + 82, + 289, + 99 + ], + "score": 1.0, + "content": "2021, virtual, 2021, pp. 14 200–14 213.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 99, + 506, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 506, + 113 + ], + "score": 1.0, + "content": "[12] J. F. Gemmeke, D. P. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 126, + 110, + 506, + 124 + ], + "score": 1.0, + "content": "M. Ritter, “Audio set: An ontology and human-labeled dataset for audio events,” in 2017 IEEE", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 121, + 507, + 135 + ], + "spans": [ + { + "bbox": [ + 126, + 121, + 507, + 135 + ], + "score": 1.0, + "content": "International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2017,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 132, + 184, + 145 + ], + "spans": [ + { + "bbox": [ + 125, + 132, + 184, + 145 + ], + "score": 1.0, + "content": "pp. 776–780.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 145, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 104, + 145, + 505, + 162 + ], + "score": 1.0, + "content": "[13] K. J. Piczak, “ESC: Dataset for Environmental Sound Classification,” in Proceedings of the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 158, + 463, + 172 + ], + "spans": [ + { + "bbox": [ + 126, + 158, + 463, + 172 + ], + "score": 1.0, + "content": "23rd Annual ACM Conference on Multimedia. ACM Press, 2015, pp. 1015–1018.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "score": 1.0, + "content": "[14] H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jégou, “Training data-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 186, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 127, + 186, + 505, + 198 + ], + "score": 1.0, + "content": "efficient image transformers & distillation through attention,” in International Conference on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 196, + 349, + 209 + ], + "spans": [ + { + "bbox": [ + 125, + 196, + 349, + 209 + ], + "score": 1.0, + "content": "Machine Learning. PMLR, 2021, pp. 10 347–10 357.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "[15] J. Carreira and A. Zisserman, “Quo vadis, action recognition? A new model and the kinetics", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 224, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 127, + 224, + 506, + 236 + ], + "score": 1.0, + "content": "dataset,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 234, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 127, + 234, + 479, + 248 + ], + "score": 1.0, + "content": "Honolulu, HI, USA, July 21-26, 2017. IEEE Computer Society, 2017, pp. 4724–4733.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 249, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 264 + ], + "score": 1.0, + "content": "[16] C. Feichtenhofer, A. Pinz, and R. P. Wildes, “Spatiotemporal residual networks for video action", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 260, + 244, + 273 + ], + "spans": [ + { + "bbox": [ + 127, + 260, + 244, + 273 + ], + "score": 1.0, + "content": "recognition,” in NIPS, 2016.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "score": 1.0, + "content": "[17] H. Bao, L. Dong, and F. Wei, “Beit: BERT pre-training of image transformers,” CoRR, vol.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 287, + 222, + 299 + ], + "spans": [ + { + "bbox": [ + 127, + 287, + 222, + 299 + ], + "score": 1.0, + "content": "abs/2106.08254, 2021.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 302, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 317 + ], + "score": 1.0, + "content": "[18] Y. Gong, C.-I. Lai, Y.-A. Chung, and J. R. Glass, “Ssast: Self-supervised audio spectrogram", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 313, + 324, + 327 + ], + "spans": [ + { + "bbox": [ + 127, + 313, + 324, + 327 + ], + "score": 1.0, + "content": "transformer,” ArXiv, vol. abs/2110.09784, 2021.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 328, + 507, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 507, + 343 + ], + "score": 1.0, + "content": "[19] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer:", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 341, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 127, + 341, + 505, + 353 + ], + "score": 1.0, + "content": "Hierarchical vision transformer using shifted windows,” in 2021 IEEE/CVF International", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 351, + 507, + 365 + ], + "spans": [ + { + "bbox": [ + 127, + 351, + 507, + 365 + ], + "score": 1.0, + "content": "Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 361, + 252, + 376 + ], + "spans": [ + { + "bbox": [ + 127, + 361, + 252, + 376 + ], + "score": 1.0, + "content": "IEEE, 2021, pp. 9992–10 002.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 376, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 506, + 391 + ], + "score": 1.0, + "content": "[20] C. Wei, H. Fan, S. Xie, C. Wu, A. L. Yuille, and C. Feichtenhofer, “Masked feature prediction", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 388, + 427, + 402 + ], + "spans": [ + { + "bbox": [ + 127, + 388, + 427, + 402 + ], + "score": 1.0, + "content": "for self-supervised visual pre-training,” CoRR, vol. abs/2112.09133, 2021.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 404, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 506, + 418 + ], + "score": 1.0, + "content": "[21] P. Vincent, H. Larochelle, Y. Bengio, and P. Manzagol, “Extracting and composing robust", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 127, + 415, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 127, + 415, + 505, + 427 + ], + "score": 1.0, + "content": "features with denoising autoencoders,” in Machine Learning, Proceedings of the Twenty-Fifth", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 426, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 127, + 426, + 506, + 440 + ], + "score": 1.0, + "content": "International Conference (ICML 2008), Helsinki, Finland, June 5-9, 2008, ser. ACM", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 438, + 468, + 450 + ], + "spans": [ + { + "bbox": [ + 127, + 438, + 468, + 450 + ], + "score": 1.0, + "content": "International Conference Proceeding Series, vol. 307. ACM, 2008, pp. 1096–1103.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 451, + 507, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 507, + 468 + ], + "score": 1.0, + "content": "[22] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P. Manzagol, “Stacked denoising", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 124, + 461, + 507, + 479 + ], + "spans": [ + { + "bbox": [ + 124, + 461, + 507, + 479 + ], + "score": 1.0, + "content": "autoencoders: Learning useful representations in a deep network with a local denoising", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 474, + 380, + 488 + ], + "spans": [ + { + "bbox": [ + 126, + 474, + 380, + 488 + ], + "score": 1.0, + "content": "criterion,” J. Mach. Learn. Res., vol. 11, pp. 3371–3408, 2010.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 490, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 505 + ], + "score": 1.0, + "content": "[23] D. Pathak, P. Krähenbühl, J. Donahue, T. Darrell, and A. A. Efros, “Context encoders: Feature", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 501, + 507, + 515 + ], + "spans": [ + { + "bbox": [ + 127, + 501, + 507, + 515 + ], + "score": 1.0, + "content": "learning by inpainting,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 509, + 508, + 528 + ], + "spans": [ + { + "bbox": [ + 126, + 509, + 508, + 528 + ], + "score": 1.0, + "content": "CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016. IEEE Computer Society, 2016, pp.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 523, + 181, + 536 + ], + "spans": [ + { + "bbox": [ + 127, + 523, + 181, + 536 + ], + "score": 1.0, + "content": "2536–2544.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 539, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 505, + 551 + ], + "score": 1.0, + "content": "[24] M. Chen, A. Radford, R. Child, J. Wu, H. Jun, D. Luan, and I. Sutskever, “Generative", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 125, + 548, + 507, + 564 + ], + "spans": [ + { + "bbox": [ + 125, + 548, + 507, + 564 + ], + "score": 1.0, + "content": "pretraining from pixels,” in Proceedings of the 37th International Conference on Machine", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 125, + 558, + 507, + 576 + ], + "spans": [ + { + "bbox": [ + 125, + 558, + 507, + 576 + ], + "score": 1.0, + "content": "Learning, ICML 2020, 13-18 July 2020, Virtual Event, ser. Proceedings of Machine Learning", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 571, + 337, + 585 + ], + "spans": [ + { + "bbox": [ + 127, + 571, + 337, + 585 + ], + "score": 1.0, + "content": "Research, vol. 119. PMLR, 2020, pp. 1691–1703.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 588, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 506, + 600 + ], + "score": 1.0, + "content": "[25] A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. Voss, A. Radford, M. Chen, and I. Sutskever,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 127, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "“Zero-shot text-to-image generation,” in Proceedings of the 38th International Conference on", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 127, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, ser. Proceedings of Machine", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 619, + 376, + 634 + ], + "spans": [ + { + "bbox": [ + 127, + 619, + 376, + 634 + ], + "score": 1.0, + "content": "Learning Research, vol. 139. PMLR, 2021, pp. 8821–8831.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "[26] N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in 2005 IEEE", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 127, + 648, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 127, + 648, + 505, + 659 + ], + "score": 1.0, + "content": "Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2005),", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 126, + 656, + 475, + 672 + ], + "spans": [ + { + "bbox": [ + 126, + 656, + 475, + 672 + ], + "score": 1.0, + "content": "20-26 June 2005, San Diego, CA, USA. IEEE Computer Society, 2005, pp. 886–893.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 673, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 506, + 687 + ], + "score": 1.0, + "content": "[27] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” 2016", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 127, + 684, + 499, + 698 + ], + "spans": [ + { + "bbox": [ + 127, + 684, + 499, + 698 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2015.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 698, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 714 + ], + "score": 1.0, + "content": "[28] K. Koutini, J. Schlüter, H. Eghbal-zadeh, and G. Widmer, “Efficient training of audio", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 125, + 709, + 385, + 725 + ], + "spans": [ + { + "bbox": [ + 125, + 709, + 385, + 725 + ], + "score": 1.0, + "content": "transformers with patchout,” CoRR, vol. abs/2110.05069, 2021.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 25.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "11", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 37, + 507, + 727 + ], + "lines": [], + "index": 25.5, + "bbox_fs": [ + 104, + 72, + 508, + 725 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 46, + 507, + 726 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "[29] K. Chen, X. Du, B. Zhu, Z. Ma, T. Berg-Kirkpatrick, and S. Dubnov, “Hts-at: A hierarchi-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 83, + 506, + 98 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 506, + 98 + ], + "score": 1.0, + "content": "cal token-semantic audio transformer for sound classification and detection,” arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 94, + 231, + 107 + ], + "spans": [ + { + "bbox": [ + 126, + 94, + 231, + 107 + ], + "score": 1.0, + "content": "arXiv:2202.00874, 2022.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 110, + 507, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 507, + 124 + ], + "score": 1.0, + "content": "[30] Y. Gong, Y. Chung, and J. Glass, “Psla: Improving audio tagging with pretraining, sampling, la-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 118, + 507, + 136 + ], + "spans": [ + { + "bbox": [ + 125, + 118, + 507, + 136 + ], + "score": 1.0, + "content": "beling, and aggregation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 129, + 156, + 146 + ], + "spans": [ + { + "bbox": [ + 126, + 129, + 156, + 146 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "score": 1.0, + "content": "[31] Y. Gong, S. Khurana, A. Rouditchenko, and J. Glass, “Cmkd: Cnn/transformer-based", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 157, + 394, + 171 + ], + "spans": [ + { + "bbox": [ + 126, + 157, + 394, + 171 + ], + "score": 1.0, + "content": "cross-model knowledge distillation for audio classification,” 2022.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 507, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 507, + 189 + ], + "score": 1.0, + "content": "[32] S. Schneider, A. Baevski, R. Collobert, and M. Auli, “wav2vec: Unsupervised pre-training", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 184, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 127, + 184, + 506, + 198 + ], + "score": 1.0, + "content": "for speech recognition,” in Interspeech 2019, 20th Annual Conference of the International", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 193, + 507, + 211 + ], + "spans": [ + { + "bbox": [ + 125, + 193, + 507, + 211 + ], + "score": 1.0, + "content": "Speech Communication Association, Graz, Austria, 15-19 September 2019. ISCA, 2019, pp.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 206, + 179, + 220 + ], + "spans": [ + { + "bbox": [ + 127, + 206, + 179, + 220 + ], + "score": 1.0, + "content": "3465–3469.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 221, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 506, + 235 + ], + "score": 1.0, + "content": "[33] A. Baevski, Y. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 232, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 127, + 232, + 506, + 246 + ], + "score": 1.0, + "content": "self-supervised learning of speech representations,” in Advances in Neural Information", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 127, + 242, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 127, + 242, + 505, + 256 + ], + "score": 1.0, + "content": "Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 254, + 338, + 268 + ], + "spans": [ + { + "bbox": [ + 126, + 254, + 338, + 268 + ], + "score": 1.0, + "content": "NeurIPS 2020, December 6-12, 2020, virtual, 2020.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 268, + 507, + 284 + ], + "spans": [ + { + "bbox": [ + 104, + 268, + 507, + 284 + ], + "score": 1.0, + "content": "[34] A. Baevski, W. Hsu, Q. Xu, A. Babu, J. Gu, and M. Auli, “data2vec: A general framework for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 125, + 280, + 497, + 295 + ], + "spans": [ + { + "bbox": [ + 125, + 280, + 497, + 295 + ], + "score": 1.0, + "content": "self-supervised learning in speech, vision and language,” CoRR, vol. abs/2202.03555, 2022.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 296, + 507, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 507, + 310 + ], + "score": 1.0, + "content": "[35] W.-N. Hsu, B. Bolte, Y.-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “Hu-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 127, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "bert: Self-supervised speech representation learning by masked prediction of hidden units,”", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 125, + 316, + 507, + 333 + ], + "spans": [ + { + "bbox": [ + 125, + 316, + 507, + 333 + ], + "score": 1.0, + "content": "IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 3451–3460,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 328, + 155, + 343 + ], + "spans": [ + { + "bbox": [ + 126, + 328, + 155, + 343 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 344, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 358 + ], + "score": 1.0, + "content": "[36] B. Shi, W. Hsu, and A. Mohamed, “Robust self-supervised audio-visual speech recognition,”", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 355, + 268, + 369 + ], + "spans": [ + { + "bbox": [ + 127, + 355, + 268, + 369 + ], + "score": 1.0, + "content": "CoRR, vol. abs/2201.01763, 2022.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 370, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 384 + ], + "score": 1.0, + "content": "[37] S. Srivastava, Y. Wang, A. Tjandra, A. Kumar, C. Liu, K. Singh, and Y. Saraf, “Conformer-based", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 381, + 504, + 396 + ], + "spans": [ + { + "bbox": [ + 127, + 381, + 504, + 396 + ], + "score": 1.0, + "content": "self-supervised learning for non-speech audio tasks,” arXiv preprint arXiv:2110.07313, 2021.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 395, + 507, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 507, + 412 + ], + "score": 1.0, + "content": "[38] A. Baade, P. Peng, and D. Harwath, “Mae-ast: Masked autoencoding audio spectrogram", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 408, + 345, + 421 + ], + "spans": [ + { + "bbox": [ + 126, + 408, + 345, + 421 + ], + "score": 1.0, + "content": "transformer,” arXiv preprint arXiv:2203.16691, 2022.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "score": 1.0, + "content": "[39] A. van den Oord, Y. Li, and O. Vinyals, “Representation learning with contrastive predictive", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 126, + 433, + 304, + 448 + ], + "spans": [ + { + "bbox": [ + 126, + 433, + 304, + 448 + ], + "score": 1.0, + "content": "coding,” CoRR, vol. abs/1807.03748, 2018.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 449, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 464 + ], + "score": 1.0, + "content": "[40] R. Arandjelovic and A. Zisserman, “Objects that sound,” in Computer Vision - ECCV 2018 -", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 127, + 460, + 507, + 475 + ], + "spans": [ + { + "bbox": [ + 127, + 460, + 507, + 475 + ], + "score": 1.0, + "content": "15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part I, ser.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 472, + 449, + 486 + ], + "spans": [ + { + "bbox": [ + 127, + 472, + 449, + 486 + ], + "score": 1.0, + "content": "Lecture Notes in Computer Science, vol. 11205. Springer, 2018, pp. 451–466.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 487, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 507, + 501 + ], + "score": 1.0, + "content": "[41] M. Patrick, P. Huang, I. Misra, F. Metze, A. Vedaldi, Y. M. Asano, and J. F. Henriques,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 497, + 506, + 512 + ], + "spans": [ + { + "bbox": [ + 126, + 497, + 506, + 512 + ], + "score": 1.0, + "content": "“Space-time crop & attend: Improving cross-modal video representation learning,” in 2021", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 125, + 507, + 507, + 524 + ], + "spans": [ + { + "bbox": [ + 125, + 507, + 507, + 524 + ], + "score": 1.0, + "content": "IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 126, + 519, + 355, + 535 + ], + "spans": [ + { + "bbox": [ + 126, + 519, + 355, + 535 + ], + "score": 1.0, + "content": "October 10-17, 2021. IEEE, 2021, pp. 10 540–10 552.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 534, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 549 + ], + "score": 1.0, + "content": "[42] A. T. Liu, S. Yang, P. Chi, P. Hsu, and H. Lee, “Mockingjay: Unsupervised speech", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 125, + 545, + 507, + 560 + ], + "spans": [ + { + "bbox": [ + 125, + 545, + 507, + 560 + ], + "score": 1.0, + "content": "representation learning with deep bidirectional transformer encoders,” in 2020 IEEE", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 556, + 507, + 571 + ], + "spans": [ + { + "bbox": [ + 127, + 556, + 507, + 571 + ], + "score": 1.0, + "content": "International Conference on Acoustics, Speech and Signal Processing, ICASSP 2020,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 568, + 389, + 582 + ], + "spans": [ + { + "bbox": [ + 127, + 568, + 389, + 582 + ], + "score": 1.0, + "content": "Barcelona, Spain, May 4-8, 2020. IEEE, 2020, pp. 6419–6423.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 583, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 597 + ], + "score": 1.0, + "content": "[43] D. Chong, H. Wang, P. Zhou, and Q. Zeng, “Masked spectrogram prediction for self-supervised", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 594, + 233, + 608 + ], + "spans": [ + { + "bbox": [ + 127, + 594, + 233, + 608 + ], + "score": 1.0, + "content": "audio pre-training,” 2022.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 609, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 506, + 624 + ], + "score": 1.0, + "content": "[44] D. Niizumi, D. Takeuchi, Y. Ohishi, N. Harada, and K. Kashino, “Masked spectrogram", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 127, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "modeling using masked autoencoders for learning general-purpose audio representation,”", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 126, + 631, + 231, + 645 + ], + "spans": [ + { + "bbox": [ + 126, + 631, + 231, + 645 + ], + "score": 1.0, + "content": "arXiv:2204.12260, 2022.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 107, + 648, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 107, + 648, + 505, + 659 + ], + "score": 1.0, + "content": "[45] Y. Li, C.-Y. Wu, H. Fan, K. Mangalam, B. Xiong, J. Malik, and C. Feichtenhofer, “Mvitv2:", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 659, + 486, + 671 + ], + "spans": [ + { + "bbox": [ + 127, + 659, + 486, + 671 + ], + "score": 1.0, + "content": "Improved multiscale vision transformers for classification and detection,” in CVPR, 2022.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 673, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 505, + 687 + ], + "score": 1.0, + "content": "[46] A. van den Oord, Y. Li, and O. Vinyals, “Representation learning with contrastive predictive", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 127, + 684, + 304, + 698 + ], + "spans": [ + { + "bbox": [ + 127, + 684, + 304, + 698 + ], + "score": 1.0, + "content": "coding,” CoRR, vol. abs/1807.03748, 2018.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 698, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 714 + ], + "score": 1.0, + "content": "[47] A. Baevski, M. Auli, and A. Mohamed, “Effectiveness of self-supervised pre-training for", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 126, + 710, + 352, + 724 + ], + "spans": [ + { + "bbox": [ + 126, + 710, + 352, + 724 + ], + "score": 1.0, + "content": "speech recognition,” CoRR, vol. abs/1911.03912, 2019.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 25.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 46, + 507, + 726 + ], + "lines": [], + "index": 25.5, + "bbox_fs": [ + 104, + 72, + 507, + 724 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 75, + 507, + 633 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 507, + 89 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 507, + 89 + ], + "score": 1.0, + "content": "[48] D. S. Park, W. Chan, Y. Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V. Le, “Specaug-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 84, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 127, + 84, + 507, + 97 + ], + "score": 1.0, + "content": "ment: A simple data augmentation method for automatic speech recognition,” ArXiv, vol.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 95, + 221, + 106 + ], + "spans": [ + { + "bbox": [ + 127, + 95, + 221, + 106 + ], + "score": 1.0, + "content": "abs/1904.08779, 2019.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "score": 1.0, + "content": "[49] J. B. Li, S. Qu, P. Huang, and F. Metze, “Audiotagging done right: 2nd comparison of deep", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 119, + 502, + 135 + ], + "spans": [ + { + "bbox": [ + 126, + 119, + 502, + 135 + ], + "score": 1.0, + "content": "learning methods for environmental sound classification,” CoRR, vol. abs/2203.13448, 2022.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 135, + 507, + 148 + ], + "spans": [ + { + "bbox": [ + 106, + 135, + 507, + 148 + ], + "score": 1.0, + "content": "[50] S. Hershey, S. Chaudhuri, D. P. W. Ellis, J. F. Gemmeke, A. Jansen, C. Moore, M. Plakal,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 146, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 126, + 146, + 506, + 160 + ], + "score": 1.0, + "content": "D. Platt, R. A. Saurous, B. Seybold, M. Slaney, R. Weiss, and K. Wilson, “Cnn architectures for", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 155, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 126, + 155, + 506, + 172 + ], + "score": 1.0, + "content": "large-scale audio classification,” in International Conference on Acoustics, Speech and Signal", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 167, + 244, + 181 + ], + "spans": [ + { + "bbox": [ + 126, + 167, + 244, + 181 + ], + "score": 1.0, + "content": "Processing (ICASSP), 2017.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 182, + 507, + 198 + ], + "spans": [ + { + "bbox": [ + 104, + 182, + 507, + 198 + ], + "score": 1.0, + "content": "[51] S. Hershey, D. P. Ellis, E. Fonseca, A. Jansen, C. Liu, R. C. Moore, and M. Plakal, “The", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 192, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 126, + 192, + 506, + 207 + ], + "score": 1.0, + "content": "benefit of temporally-strong labels in audio event classification,” in ICASSP 2021-2021 IEEE", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 125, + 203, + 507, + 219 + ], + "spans": [ + { + "bbox": [ + 125, + 203, + 507, + 219 + ], + "score": 1.0, + "content": "International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 216, + 183, + 228 + ], + "spans": [ + { + "bbox": [ + 126, + 216, + 183, + 228 + ], + "score": 1.0, + "content": "pp. 366–370.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "[52] P. Warden, “Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition,” ArXiv", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 126, + 242, + 209, + 254 + ], + "spans": [ + { + "bbox": [ + 126, + 242, + 209, + 254 + ], + "score": 1.0, + "content": "e-prints, Apr. 2018.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 255, + 507, + 271 + ], + "spans": [ + { + "bbox": [ + 104, + 255, + 507, + 271 + ], + "score": 1.0, + "content": "[53] S. wen Yang, P.-H. Chi, Y.-S. Chuang, C.-I. J. Lai, K. Lakhotia, Y. Y. Lin, A. T. Liu, J. Shi,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 124, + 266, + 507, + 282 + ], + "spans": [ + { + "bbox": [ + 124, + 266, + 507, + 282 + ], + "score": 1.0, + "content": "X. Chang, G.-T. Lin, T.-H. Huang, W.-C. Tseng, K. tik Lee, D.-R. Liu, Z. Huang, S. Dong,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 127, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "S.-W. Li, S. Watanabe, A. Mohamed, and H. yi Lee, “SUPERB: Speech Processing Universal", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 288, + 438, + 302 + ], + "spans": [ + { + "bbox": [ + 126, + 288, + 438, + 302 + ], + "score": 1.0, + "content": "PERformance Benchmark,” in Proc. Interspeech 2021, 2021, pp. 1194–1198.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "[54] A. Nagrani, J. S. Chung, W. Xie, and A. Zisserman, “Voxceleb: Large-scale speaker verification", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 316, + 331, + 327 + ], + "spans": [ + { + "bbox": [ + 127, + 316, + 331, + 327 + ], + "score": 1.0, + "content": "in the wild,” Comput. Speech Lang., vol. 60, 2020.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 328, + 507, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 507, + 344 + ], + "score": 1.0, + "content": "[55] D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 127, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "P. Motlicek, Y. Qian, P. Schwarz et al., “The kaldi speech recognition toolkit,” in IEEE 2011", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 350, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 126, + 350, + 506, + 365 + ], + "score": 1.0, + "content": "workshop on automatic speech recognition and understanding, no. CONF. IEEE Signal", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 362, + 234, + 376 + ], + "spans": [ + { + "bbox": [ + 127, + 362, + 234, + 376 + ], + "score": 1.0, + "content": "Processing Society, 2011.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 375, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 104, + 375, + 506, + 393 + ], + "score": 1.0, + "content": "[56] S. Yun, D. Han, S. Chun, S. J. Oh, Y. Yoo, and J. Choe, “Cutmix: Regularization strategy to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 127, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "train strong classifiers with localizable features,” in 2019 IEEE/CVF International Conference", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 398, + 507, + 413 + ], + "spans": [ + { + "bbox": [ + 126, + 398, + 507, + 413 + ], + "score": 1.0, + "content": "on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 409, + 245, + 424 + ], + "spans": [ + { + "bbox": [ + 127, + 409, + 245, + 424 + ], + "score": 1.0, + "content": "IEEE, 2019, pp. 6022–6031.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 424, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 439 + ], + "score": 1.0, + "content": "[57] H. Zhang, M. Cissé, Y. N. Dauphin, and D. Lopez-Paz, “mixup: Beyond empirical", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 127, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 127, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "risk minimization,” in 6th International Conference on Learning Representations, ICLR", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 126, + 445, + 507, + 462 + ], + "spans": [ + { + "bbox": [ + 126, + 445, + 507, + 462 + ], + "score": 1.0, + "content": "2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 458, + 225, + 471 + ], + "spans": [ + { + "bbox": [ + 127, + 458, + 225, + 471 + ], + "score": 1.0, + "content": "OpenReview.net, 2018.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 471, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 104, + 471, + 506, + 486 + ], + "score": 1.0, + "content": "[58] S. Verbitskiy, V. Berikov, and V. Vyshegorodtsev, “Eranns: Efficient residual audio neural", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 484, + 453, + 497 + ], + "spans": [ + { + "bbox": [ + 126, + 484, + 453, + 497 + ], + "score": 1.0, + "content": "networks for audio pattern recognition,” arXiv preprint arXiv:2106.01621, 2021.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "[59] Q. Kong, Y. Cao, T. Iqbal, Y. Wang, W. Wang, and M. D. Plumbley, “Panns: Large-scale", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 126, + 509, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 126, + 509, + 506, + 523 + ], + "score": 1.0, + "content": "pretrained audio neural networks for audio pattern recognition,” IEEE ACM Trans. Audio", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 520, + 345, + 534 + ], + "spans": [ + { + "bbox": [ + 127, + 520, + 345, + 534 + ], + "score": 1.0, + "content": "Speech Lang. Process., vol. 28, pp. 2880–2894, 2020.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 534, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 104, + 534, + 506, + 549 + ], + "score": 1.0, + "content": "[60] M. Tan and Q. V. Le, “Efficientnet: Rethinking model scaling for convolutional neural", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 127, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "networks,” in Proceedings of the 36th International Conference on Machine Learning, ICML", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 126, + 555, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 126, + 555, + 506, + 572 + ], + "score": 1.0, + "content": "2019, 9-15 June 2019, Long Beach, California, USA, ser. Proceedings of Machine Learning", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 567, + 331, + 581 + ], + "spans": [ + { + "bbox": [ + 127, + 567, + 331, + 581 + ], + "score": 1.0, + "content": "Research, vol. 97. PMLR, 2019, pp. 6105–6114.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "[61] D. Griffin and J. Lim, “Signal estimation from modified short-time fourier transform,” IEEE", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 594, + 502, + 608 + ], + "spans": [ + { + "bbox": [ + 127, + 594, + 502, + 608 + ], + "score": 1.0, + "content": "Transactions on Acoustics, Speech, and Signal Processing, vol. 32, no. 2, pp. 236–243, 1984.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "[62] Y. Suzuki and H. Takeshima, “Equal-loudness-level contours for pure tones,” The Journal of the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 126, + 619, + 395, + 633 + ], + "spans": [ + { + "bbox": [ + 126, + 619, + 395, + 633 + ], + "score": 1.0, + "content": "Acoustical Society of America, vol. 116, no. 2, pp. 918–933, 2004.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 22.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 75, + 507, + 633 + ], + "lines": [], + "index": 22.5, + "bbox_fs": [ + 104, + 70, + 507, + 633 + ], + "lines_deleted": true + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file