thangquang09 commited on
Commit
d99ffd8
·
verified ·
1 Parent(s): 9bc8ac6

Release ViInstruct v1: annotation layer, README and license

Browse files
Files changed (6) hide show
  1. .gitattributes +3 -0
  2. LICENSE.md +20 -0
  3. README.md +184 -0
  4. data/test.csv +3 -0
  5. data/train.csv +3 -0
  6. data/validation.csv +3 -0
.gitattributes CHANGED
@@ -58,3 +58,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
 
 
 
 
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
61
+ data/test.csv filter=lfs diff=lfs merge=lfs -text
62
+ data/train.csv filter=lfs diff=lfs merge=lfs -text
63
+ data/validation.csv filter=lfs diff=lfs merge=lfs -text
LICENSE.md ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ViInstruct license
2
+
3
+ ViInstruct is an annotation layer: generated instructions, attribute labels, transcripts and split definitions keyed to identifiers of six source corpora. It contains no audio.
4
+
5
+ 1. **Non-commercial research use only.** You may use, copy and adapt the annotation layer for non-commercial research and teaching. Commercial use is not permitted.
6
+ 2. **Upstream terms still apply.** Every row belongs to one source corpus (column `dataset`) and remains subject to that corpus's terms:
7
+
8
+ | `dataset` | Source corpus | Upstream terms |
9
+ |---|---|---|
10
+ | `vivoice` | viVoice | CC BY-NC-SA 4.0 (gated) |
11
+ | `dolly` | Dolly-Audio | CC BY-NC-SA 4.0 (gated) |
12
+ | `vimd` | ViMD | CC BY-NC-ND 4.0 |
13
+ | `visec` | ViSEC | CC BY 4.0 |
14
+ | `lsvsc` | LSVSC | no license published |
15
+ | `vnemos` | VNEMOS | no license published |
16
+
17
+ Where the upstream terms are stricter than point 1, the upstream terms win. For LSVSC and VNEMOS, obtain permission from the original authors before any use beyond research.
18
+ 3. **Audio.** Obtain audio only from the original corpora, under their own terms.
19
+ 4. **Attribution.** Cite the ViInstruct paper and the source corpora you use.
20
+ 5. **No warranty.** The data is provided as is. Gender, age and accent labels may be model predictions and are not claims about any individual.
README.md ADDED
@@ -0,0 +1,184 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: ViInstruct
3
+ language:
4
+ - vi
5
+ license: other
6
+ license_name: viinstruct-noncommercial
7
+ license_link: LICENSE.md
8
+ task_categories:
9
+ - text-to-speech
10
+ tags:
11
+ - instruction-based-tts
12
+ - controllable-tts
13
+ - speech-style-captioning
14
+ - vietnamese
15
+ - accent
16
+ - prosody
17
+ size_categories:
18
+ - 1M<n<10M
19
+ configs:
20
+ - config_name: default
21
+ data_files:
22
+ - split: train
23
+ path: data/train.csv
24
+ - split: validation
25
+ path: data/validation.csv
26
+ - split: test
27
+ path: data/test.csv
28
+ ---
29
+
30
+ # ViInstruct
31
+
32
+ ViInstruct is a large-scale Vietnamese dataset for **instruction-based text-to-speech (TTS)**. Each utterance is paired with one free-form Vietnamese instruction that describes how it is spoken: speaker gender, age group and regional accent, plus speaking rate, pitch, loudness, expressiveness and (for the emotion subset) emotion.
33
+
34
+ - **1,224,945** utterance–instruction pairs, about **1,718 hours** of speech
35
+ - **6** source corpora covering **3** tasks: general speech, regional accent, emotion
36
+ - Instructions written by an LLM (Qwen3.5-122B-A10B) from acoustic and speaker attributes, average **22.1 words**
37
+ - Paper: *ViInstruct: A Large-Scale Vietnamese Dataset for Instruction-Based Text-to-Speech Synthesis*, APSIPA ASC 2026
38
+ - Demo with audio samples: https://thangquang09.github.io/viinstruct-demo/
39
+
40
+ > **This repository is an annotation layer, not an audio corpus.** It holds the instructions, attribute labels, transcripts and split definitions, keyed to the identifiers of the source corpora. Get the audio from each original corpus (see [Source corpora](#source-corpora)) and join on `dataset` + `source_id` / `audio_path`.
41
+
42
+ ## Example
43
+
44
+ | Field | Value |
45
+ |---|---|
46
+ | `caption_full` | Giọng nữ trưởng thành miền Nam, nói nhanh, to, giọng cao và đều đều. |
47
+ | English gloss | *Adult female voice from the South, fast, loud, high-pitched and monotone.* |
48
+ | `kw_gender` / `kw_age` / `region` | nữ / người trưởng thành / south |
49
+ | `kw_speaking_rate` / `kw_pitch` / `kw_loudness_lufs` / `kw_pitch_cv` | nhanh / cao / to / đều đều |
50
+
51
+ ```python
52
+ from datasets import load_dataset
53
+
54
+ ds = load_dataset("thangquang09/ViInstruct")
55
+ print(ds["train"][0]["caption_full"])
56
+ ```
57
+
58
+ ## Dataset structure
59
+
60
+ ### Splits
61
+
62
+ | Source | Task | Train | Validation | Test | Total |
63
+ |---|---|---:|---:|---:|---:|
64
+ | Vivoice | general | 523,169 | 29,065 | 29,065 | 581,299 |
65
+ | Dolly | general | 526,067 | 29,210 | 29,232 | 584,509 |
66
+ | ViMD | accent | 10,087 | 1,293 | 1,410 | 12,790 |
67
+ | LSVSC | accent | 37,181 | 3,800 | 3,008 | 43,989 |
68
+ | VNEMOS | emotion | 189 | 11 | 11 | 211 |
69
+ | ViSEC | emotion | 1,932 | 107 | 108 | 2,147 |
70
+ | **Total** | | **1,098,625** | **63,486** | **62,834** | **1,224,945** |
71
+
72
+ ViMD and LSVSC keep their official splits. Dolly is split speaker-disjointly by `voice_id`. Vivoice, VNEMOS and ViSEC have no official split or stable speaker id, so they use a stratified 90/5/5 split at the sample level.
73
+
74
+ ### Fields
75
+
76
+ All descriptor fields hold the Vietnamese keyword that was given to the LLM. An empty string means the attribute does not apply (for example `kw_emotion` outside the emotion subset).
77
+
78
+ | Column | Description | Values |
79
+ |---|---|---|
80
+ | `id` | Unique id, `{dataset}_{source_id}` | |
81
+ | `source_id` | Utterance id in the source corpus | |
82
+ | `dataset` | Source corpus | `vivoice`, `dolly`, `vimd`, `lsvsc`, `vnemos`, `visec` |
83
+ | `task` | Subset | `general`, `accent`, `emotion` |
84
+ | `split` | Split | `train`, `val`, `test` |
85
+ | `transcript` | Text of the utterance | |
86
+ | `caption_full` | The natural-language instruction | |
87
+ | `kw_gender` | Gender | `nam` (male), `nữ` (female) |
88
+ | `kw_age` | Age group | `trẻ em` (child), `thanh thiếu niên` (teen), `người trưởng thành` (adult), `cao tuổi` (senior) |
89
+ | `kw_accent` | Regional accent | `giọng miền Bắc` / `giọng miền Trung` / `giọng miền Nam` |
90
+ | `kw_accent_province` | Province (ViMD only, 63 provinces) | |
91
+ | `region` | Machine-readable accent region | `north`, `central`, `south` |
92
+ | `kw_speaking_rate` | Speaking rate, 7 levels | `rất chậm`, `chậm`, `hơi chậm`, `vừa phải`, `hơi nhanh`, `nhanh`, `rất nhanh` |
93
+ | `kw_pitch` | Mean pitch, 7 levels (binned per gender–accent group) | `rất trầm`, `trầm`, `hơi trầm`, `trung bình`, `hơi cao`, `cao`, `rất cao` |
94
+ | `kw_loudness_lufs` | Loudness, 5 levels | `rất nhỏ`, `nhỏ`, `vừa`, `to`, `rất to` |
95
+ | `kw_pitch_cv` | Pitch variation (expressiveness), 5 levels | `đều đều`, `ít biểu cảm`, `biểu cảm vừa phải`, `sinh động`, `rất biểu cảm` |
96
+ | `kw_emotion` | Emotion (emotion subset only) | `trung lập` (neutral), `vui vẻ` (happiness), `buồn bã` (sadness), `tức giận` (anger), `sợ hãi` (fear) |
97
+ | `audio_path` | Path of the audio file relative to the root of its source corpus | |
98
+
99
+ ### Attribute distribution
100
+
101
+ - Gender: 53.8% male, 46.2% female
102
+ - Age: 96.6% adult, 1.8% children, 1.1% teens, 0.4% seniors
103
+ - Accent: 59.6% North, 23.9% Central, 16.5% South
104
+ - Emotion (2,358 clips): anger, fear, happiness, neutral, sadness. Fear comes only from VNEMOS, whose *anxiety* class is mapped to fear.
105
+ - Median duration 4.31 s; mean pitch about 137 Hz (male) and 231 Hz (female); speaking rate about 4.57 syllables/s; loudness about −21.4 LUFS.
106
+
107
+ ## Source corpora
108
+
109
+ | Corpus | Task | Content | Kept clips | Kept hours | Upstream license |
110
+ |---|---|---|---:|---:|---|
111
+ | [viVoice](https://hf.co/datasets/capleaf/viVoice) | general | read and spontaneous speech from public video channels | 581,299 | 779 | CC BY-NC-SA 4.0 (gated) |
112
+ | [Dolly-Audio](https://hf.co/datasets/dolly-vn/dolly-audio-1000h-vietnamese) | general | TTS-synthesized speech, 484 voice presets | 584,509 | 808 | CC BY-NC-SA 4.0 (gated) |
113
+ | ViMD (Dinh et al., EMNLP 2024) | accent | read speech, 63 provinces | 12,790 | 54 | CC BY-NC-ND 4.0 |
114
+ | LSVSC (Tran et al., Electronics 2024) | accent | read speech with gender / age / region labels | 43,989 | 75 | no license published |
115
+ | VNEMOS (Quang Anh et al., ICDV 2024) | emotion | film and TV recordings | 211 | 0.35 | no license published |
116
+ | ViSEC (Thanh et al., ICASSP 2024) | emotion | acted speech from public video | 2,147 | 1.12 | CC BY 4.0 |
117
+
118
+ 1,583,772 raw clips (2,223 h) were reduced to 1,224,945 (1,718 h) by quality filtering and deduplication. Dolly is synthetic and makes up 47.7% of the pairs.
119
+
120
+ ## How the dataset was built
121
+
122
+ 1. **Acoustic features and quality filtering.** Duration, speaking rate (syllables per second from the transcript), pitch statistics with CREPE, loudness in LUFS, and speech proportion with Silero VAD. Clips are kept when duration ≥ 2.0 s (≥ 1.0 s for ViSEC), SNR ≥ 15 dB, clipping ratio < 0.01 and speech proportion ≥ 0.30.
123
+ 2. **Speaker attributes.** Ground-truth labels are kept whenever the source corpus provides them. Otherwise a wav2vec2 gender classifier, a ViMD-trained wav2vec2 accent classifier (94.6% macro-F1) and a fine-tuned MERaLiON age classifier (teens / adults / seniors) predict them. ViMD transcripts come from PhoWhisper.
124
+ 3. **Quantile binning.** Continuous features become Vietnamese keywords with quantile bins: 7 levels for pitch and speaking rate, 5 for loudness and pitch variation. Pitch is binned per gender–accent group. Bin edges are computed on Vivoice and reused for the other corpora.
125
+ 4. **Instruction generation.** Qwen3.5-122B-A10B writes one instruction per utterance from the transcript, the shuffled attribute tags and three few-shot examples. Synonym substitution, random attribute order, varied few-shot examples and a per-utterance temperature in [0.3, 0.7] increase lexical variety.
126
+
127
+ ### Instruction quality audit
128
+
129
+ All 1,224,945 instructions were matched against their own tags with a Vietnamese descriptor lexicon (text only, no listening):
130
+
131
+ - 81.4% of instructions realize every tag in words; the rest paraphrase outside the lexicon.
132
+ - When a slot is realized, it names the conditioning level at least 96.6% of the time (loudness is the lowest; gender and accent are ≥ 99.99%).
133
+ - 99.7% of disagreements are one bin apart.
134
+ - No emotion word appears in the 1,222,587 instructions without an emotion tag.
135
+ - Known defect: 4.7% of instructions render *người trưởng thành* (adult) as *lớn tuổi* (elderly).
136
+
137
+ This audit checks tag → text only. Whether the tags match the audio rests on the ground-truth labels and the classifiers above.
138
+
139
+ ## Benchmark
140
+
141
+ The paper adapts two baselines to Vietnamese (Parler-TTS and CapSpeech NAR, both with a frozen ViT5-large instruction encoder) and evaluates them on a general full-tag set (1,228 clips) and an accent set (1,710 clips).
142
+
143
+ | Metric | Reference audio | Parler-TTS | CapSpeech NAR |
144
+ |---|---:|---:|---:|
145
+ | CER (%) ↓ | 0.96 | 3.19 | 0.85 |
146
+ | WER (%) ↓ | 2.15 | 4.80 | 1.55 |
147
+ | UTMOSv2 ↑ | 2.95 | 2.54 | 2.65 |
148
+ | Gender macro-F1 (%) ↑ | | 84.83 | 100.00 |
149
+ | Age macro-F1 (%) ↑ | | 49.94 | 52.82 |
150
+ | Accent macro-F1, general set (%) ↑ | | 48.56 | 77.48 |
151
+ | Accent macro-F1, accent set (%) ↑ | | 38.82 | 72.37 |
152
+ | Acoustic bin accuracy (%) ↑ | | 36.46 | 66.55 |
153
+
154
+ Chance level is about 33% macro-F1 for the three-way age and accent tasks and 14–20% for bin accuracy.
155
+
156
+ ## Limitations
157
+
158
+ - Age labels are heavily adult and partly come from a coarse three-class classifier; age adherence of both baselines is close to chance.
159
+ - The emotion subset is small (2,358 clips, 1.5 h) and has no reliable automatic metric.
160
+ - The accent subset is regionally imbalanced toward Northern speech.
161
+ - Pitch is summarized by scalar statistics that do not separate lexical tone from paralinguistic prosody.
162
+ - Nearly half of the pairs come from synthesized speech (Dolly).
163
+ - Predicted attributes and the benchmark are automatic. No human listening study has been done yet.
164
+
165
+ ## License
166
+
167
+ The annotation layer (instructions, attribute labels and splits) is released for **non-commercial research use only**. Each subset also remains governed by the terms of its source corpus, listed in [Source corpora](#source-corpora). The upstream terms are mutually incompatible or unstated, so this repository does not redistribute any audio. See [LICENSE.md](LICENSE.md).
168
+
169
+ ## Ethics
170
+
171
+ No new speech was recorded and no attempt was made to identify speakers. Where ground-truth labels are absent, gender, age and accent are model predictions, not self-reported identity, and should not be read as factual claims about individuals. The instructions inherit that uncertainty.
172
+
173
+ ## Citation
174
+
175
+ ```bibtex
176
+ @inproceedings{ly2026viinstruct,
177
+ title = {{ViInstruct}: A Large-Scale {Vietnamese} Dataset for Instruction-Based Text-to-Speech Synthesis},
178
+ author = {Ly, Thang and Do, Tri-Nhan and Nguyen, Tu-Anh and Mac, Dang-Khoa},
179
+ booktitle = {Proc. Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC)},
180
+ year = {2026}
181
+ }
182
+ ```
183
+
184
+ Please also cite the source corpora you use.
data/test.csv ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b6e1858ee177c26cde7e4630a265d10355360461d03a1642906b81013b1d0c41
3
+ size 31930977
data/train.csv ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d3bc0a3c0a75121e17a86dc8c517231cd53c4f344a14699004f55ee8dcbe3c2c
3
+ size 572370506
data/validation.csv ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:deb5a5da10aa56aec4982624b0f585ade85c0d8f427a46b2e80e46b026bf1ac6
3
+ size 32162402