zyk21 commited on
Commit
373f052
·
verified ·
1 Parent(s): 5d5f955

Improve README structure and DubBench documentation

Browse files
Files changed (1) hide show
  1. README.md +208 -163
README.md CHANGED
@@ -1,163 +1,208 @@
1
- ---
2
- license: cc-by-nc-4.0
3
- task_categories:
4
- - text-to-speech
5
- language:
6
- - zh
7
- - en
8
- tags:
9
- - audio-visual
10
- - video-dubbing
11
- - lip-sync
12
- ---
13
-
14
- # VoxDub
15
-
16
- A bilingual lip-synchronized audio-visual dataset for automatic video dubbing.
17
-
18
- - **Paper**: [VoxDub: A Bilingual Lip-Synchronized Audio-Visual Dataset for Automatic Video Dubbing](https://huggingface.co/papers/2610.25259)
19
- - **Project page**: [https://voxdub-dataset.github.io/VoxDub/](https://voxdub-dataset.github.io/VoxDub/)
20
- - **GitHub repository**: [https://github.com/voxdub-dataset/VoxDub](https://github.com/voxdub-dataset/VoxDub)
21
-
22
- ## Dataset Summary
23
-
24
- VoxDub contains approximately **766K clips** and **1.58K hours** of Chinese and English audio-visual recordings collected from in-the-wild videos. It is designed for automatic video dubbing (AVD), with downstream evaluations on both text-to-speech (TTS) and AVD.
25
-
26
- The curation pipeline combines speech-centered segmentation with dual-ASR transcription checking, speech quality filtering, face visibility checks, and lip-speech synchronization verification. The public data release contains **annotations**; source video and audio files are retrieved separately using the provided reconstruction tools.
27
-
28
- ## Quick Start
29
-
30
- After installation, download the annotations, extract them, and reconstruct a small subset:
31
-
32
- ```bash
33
- hf download zyk21/VoxDub av_segments_v1.tar.zst --repo-type dataset --local-dir annotations
34
- python scripts/pipeline.py extract
35
- python scripts/pipeline.py run --from-annots --limit 10 --device cpu
36
- ```
37
-
38
- For a configured CUDA environment, replace `--device cpu` with `--device cuda`.
39
-
40
- `--limit 10` selects up to **10 source videos**, processing their annotated segments. It does not limit the output to 10 clips.
41
-
42
- See the [GitHub README](https://github.com/voxdub-dataset/VoxDub) for full usage details.
43
-
44
- ## DubBench
45
-
46
- DubBench contains English and Mandarin Chinese subsets. Each sample pairs a target utterance with a reference utterance.
47
-
48
- In our experiments, the reference audio and text are used as conditioning inputs, while the target text specifies the content to generate. Only AVD additionally uses the target video. The generated audio is evaluated against the target audio.
49
-
50
- ### Data sources
51
-
52
- - **English:** [CelebV-Dub](https://voicecraft-dub.github.io/)
53
- - **Chinese:** [CMLR](https://www.vipazoo.cn/CMLR.html)
54
-
55
- Please refer to the original datasets for their papers, licenses, and usage terms.
56
-
57
- ### Directory structure
58
-
59
- ```text
60
- dubbench/
61
- ├── english/
62
- │ ├── metadata.tsv
63
- │ ├── target.tar
64
- │ └── reference.tar
65
- └── chinese/
66
- ├── metadata.tsv
67
- ├── target.tar
68
- └── reference.tar
69
- ```
70
-
71
- The four TAR archives are uncompressed. After extraction, each language subset has the following structure:
72
-
73
- ```text
74
- english/ or chinese/
75
- ├── metadata.tsv
76
- ├── target/
77
- │ ├── <utterance_id>.mp4
78
- │ ├── <utterance_id>.wav
79
- │ └── <utterance_id>.txt
80
- └── reference/
81
- ├── <utterance_id>.wav
82
- └── <utterance_id>.txt
83
- ```
84
-
85
- Each target utterance includes `.mp4`, `.wav`, and `.txt` files, while each reference utterance includes `.wav` and `.txt` files.
86
-
87
- ### Metadata format
88
-
89
- Each `metadata.tsv` file is UTF-8 encoded, has no header row, and contains four tab-separated columns:
90
-
91
- | Column | Field | Description |
92
- |---:|---|---|
93
- | 1 | `target_id` | Extensionless path of the target utterance |
94
- | 2 | `target_text` | Text to be generated |
95
- | 3 | `reference_id` | Extensionless path of the reference utterance |
96
- | 4 | `reference_text` | Text associated with the reference utterance |
97
-
98
- Example:
99
-
100
- ```text
101
- target/-4GNkXEUOwY_6_0_3<TAB>THERE'S NOTHING TO INSTALL...<TAB>reference/-4GNkXEUOwY_0_1_0<TAB>AND WHILE IT IS EASY...
102
- ```
103
-
104
- Resolve an asset by joining the language subset directory with its utterance ID and file extension:
105
-
106
- ```text
107
- dubbench/english/target/-4GNkXEUOwY_6_0_3.wav
108
- ```
109
-
110
- ### Python example
111
-
112
- ```python
113
- import csv
114
- import tarfile
115
- from pathlib import Path
116
-
117
- from huggingface_hub import snapshot_download
118
-
119
-
120
- root = Path(
121
- snapshot_download(
122
- repo_id="zyk21/VoxDub",
123
- repo_type="dataset",
124
- allow_patterns="dubbench/**",
125
- )
126
- )
127
-
128
- subset = root / "dubbench" / "english" # or "chinese"
129
-
130
- for role in ("target", "reference"):
131
- with tarfile.open(subset / f"{role}.tar") as archive:
132
- archive.extractall(subset)
133
-
134
- with (subset / "metadata.tsv").open(encoding="utf-8", newline="") as file:
135
- rows = csv.reader(file, delimiter="\t")
136
-
137
- for target_id, target_text, reference_id, reference_text in rows:
138
- target_audio = subset / f"{target_id}.wav"
139
- target_video = subset / f"{target_id}.mp4"
140
- reference_audio = subset / f"{reference_id}.wav"
141
-
142
- assert target_audio.is_file()
143
- assert target_video.is_file()
144
- assert reference_audio.is_file()
145
-
146
- # All systems use reference_audio, reference_text, and target_text.
147
- # AVD additionally uses target_video.
148
- # The generated audio is evaluated against target_audio.
149
- ```
150
-
151
- Reference utterances may be reused by multiple metadata rows. Samples should therefore be resolved through the IDs in `metadata.tsv` rather than by directory order.
152
-
153
- ### Attribution and terms
154
-
155
- The English and Chinese subsets are derived from CelebV-Dub and CMLR, respectively. Please cite the corresponding datasets and publications when using DubBench.
156
-
157
- Users must follow the licenses and usage terms of the original datasets and source media. This repository does not grant additional rights to the original audio or video.
158
-
159
- ## License
160
-
161
- - **Repository code:** MIT License.
162
- - **Dataset annotations:** [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
163
- - **Source media:** rights remain with the original rights holders; the annotation release does not grant rights to the source videos or audio.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-nc-4.0
3
+ task_categories:
4
+ - text-to-speech
5
+ language:
6
+ - zh
7
+ - en
8
+ tags:
9
+ - audio-visual
10
+ - video-dubbing
11
+ - lip-sync
12
+ ---
13
+
14
+ # VoxDub
15
+
16
+ A bilingual lip-synchronized audio-visual dataset for automatic video dubbing.
17
+
18
+ - **Paper**: [VoxDub: A Bilingual Lip-Synchronized Audio-Visual Dataset for Automatic Video Dubbing](https://huggingface.co/papers/2610.25259)
19
+ - **Project page**: [https://voxdub-dataset.github.io/VoxDub/](https://voxdub-dataset.github.io/VoxDub/)
20
+ - **GitHub repository**: [https://github.com/voxdub-dataset/VoxDub](https://github.com/voxdub-dataset/VoxDub)
21
+
22
+ ## Dataset Summary
23
+
24
+ VoxDub contains approximately **766K clips** and **1.58K hours** of Chinese and English audio-visual recordings collected from in-the-wild videos. It is designed for automatic video dubbing (AVD), with downstream evaluations on both text-to-speech (TTS) and AVD.
25
+
26
+ The curation pipeline combines speech-centered segmentation with dual-ASR transcription checking, speech quality filtering, face visibility checks, and lip-speech synchronization verification. The public data release contains **annotations**; source video and audio files are retrieved separately using the provided reconstruction tools.
27
+
28
+ ## Quick Start
29
+
30
+ After installation, download the annotations, extract them, and reconstruct a small subset:
31
+
32
+ ```bash
33
+ hf download zyk21/VoxDub av_segments_v1.tar.zst \
34
+ --repo-type dataset \
35
+ --local-dir annotations
36
+
37
+ python scripts/pipeline.py extract
38
+ python scripts/pipeline.py run --from-annots --limit 10 --device cpu
39
+ ```
40
+
41
+ For a configured CUDA environment, replace `--device cpu` with `--device cuda`.
42
+
43
+ `--limit 10` selects up to **10 source videos**, processing their annotated segments. It does not limit the output to 10 clips.
44
+
45
+ See the [GitHub README](https://github.com/voxdub-dataset/VoxDub) for full usage details.
46
+
47
+ ## DubBench
48
+
49
+ ### 1. Overview
50
+
51
+ DubBench contains English and Mandarin Chinese subsets. Each sample pairs a target utterance with a reference utterance.
52
+
53
+ In our experiments:
54
+
55
+ - **Inputs for all systems:** reference audio, reference text, and target text
56
+ - **Additional input for AVD:** target video
57
+ - **Output:** generated target audio
58
+ - **Evaluation reference:** ground-truth target audio
59
+
60
+ ### 2. Data Sources
61
+
62
+ - **English:** [CelebV-Dub](https://voicecraft-dub.github.io/)
63
+ - **Chinese:** [CMLR](https://www.vipazoo.cn/CMLR.html)
64
+
65
+ Please refer to the original datasets for their papers, licenses, and usage terms.
66
+
67
+ ### 3. Dataset Structure
68
+
69
+ #### 3.1 Repository Layout
70
+
71
+ ```text
72
+ dubbench/
73
+ ├── english/
74
+ │ ├── metadata.tsv
75
+ │ ├── target.tar
76
+ │ └── reference.tar
77
+ └── chinese/
78
+ ├── metadata.tsv
79
+ ├── target.tar
80
+ └── reference.tar
81
+ ```
82
+
83
+ The four TAR archives are uncompressed.
84
+
85
+ #### 3.2 Extracted Files
86
+
87
+ After extraction, each language subset has the following structure:
88
+
89
+ ```text
90
+ english/ or chinese/
91
+ ├── metadata.tsv
92
+ ├── target/
93
+ │ ├── <utterance_id>.mp4
94
+ │ ├── <utterance_id>.wav
95
+ │ └── <utterance_id>.txt
96
+ └── reference/
97
+ ├── <utterance_id>.wav
98
+ └── <utterance_id>.txt
99
+ ```
100
+
101
+ Each target utterance includes `.mp4`, `.wav`, and `.txt` files. Each reference utterance includes `.wav` and `.txt` files.
102
+
103
+ ### 4. Metadata Format
104
+
105
+ Each `metadata.tsv` file is UTF-8 encoded, has no header row, and contains four tab-separated columns:
106
+
107
+ | Column | Field | Description |
108
+ |---:|---|---|
109
+ | 1 | `target_id` | Extensionless path of the target utterance |
110
+ | 2 | `target_text` | Text to be generated |
111
+ | 3 | `reference_id` | Extensionless path of the reference utterance |
112
+ | 4 | `reference_text` | Text associated with the reference utterance |
113
+
114
+ Example:
115
+
116
+ ```text
117
+ target/-4GNkXEUOwY_6_0_3<TAB>THERE'S NOTHING TO INSTALL...<TAB>reference/-4GNkXEUOwY_0_1_0<TAB>AND WHILE IT IS EASY...
118
+ ```
119
+
120
+ Resolve an asset by joining the language subset directory with its utterance ID and file extension:
121
+
122
+ ```text
123
+ dubbench/english/target/-4GNkXEUOwY_6_0_3.wav
124
+ ```
125
+
126
+ Reference utterances may be reused by multiple metadata rows. Samples should therefore be resolved through the IDs in `metadata.tsv` rather than by directory order.
127
+
128
+ ### 5. Usage
129
+
130
+ #### 5.1 Download
131
+
132
+ Download DubBench with the Hugging Face CLI:
133
+
134
+ ```bash
135
+ hf download zyk21/VoxDub \
136
+ --repo-type dataset \
137
+ --include "dubbench/**" \
138
+ --local-dir VoxDub
139
+ ```
140
+
141
+ #### 5.2 Extract
142
+
143
+ Extract the English subset:
144
+
145
+ ```bash
146
+ tar -xf VoxDub/dubbench/english/target.tar \
147
+ -C VoxDub/dubbench/english
148
+
149
+ tar -xf VoxDub/dubbench/english/reference.tar \
150
+ -C VoxDub/dubbench/english
151
+ ```
152
+
153
+ Replace `english` with `chinese` to extract the Chinese subset.
154
+
155
+ #### 5.3 Python Example
156
+
157
+ ```python
158
+ import csv
159
+ import tarfile
160
+ from pathlib import Path
161
+
162
+ from huggingface_hub import snapshot_download
163
+
164
+
165
+ root = Path(
166
+ snapshot_download(
167
+ repo_id="zyk21/VoxDub",
168
+ repo_type="dataset",
169
+ allow_patterns="dubbench/**",
170
+ )
171
+ )
172
+
173
+ subset = root / "dubbench" / "english" # or "chinese"
174
+
175
+ for role in ("target", "reference"):
176
+ with tarfile.open(subset / f"{role}.tar") as archive:
177
+ archive.extractall(subset)
178
+
179
+ with (subset / "metadata.tsv").open(encoding="utf-8", newline="") as file:
180
+ rows = csv.reader(file, delimiter="\t")
181
+
182
+ for target_id, target_text, reference_id, reference_text in rows:
183
+ target_audio = subset / f"{target_id}.wav"
184
+ target_video = subset / f"{target_id}.mp4"
185
+ reference_audio = subset / f"{reference_id}.wav"
186
+
187
+ assert target_audio.is_file()
188
+ assert target_video.is_file()
189
+ assert reference_audio.is_file()
190
+
191
+ # All systems use:
192
+ # reference_audio, reference_text, and target_text
193
+ #
194
+ # AVD additionally uses target_video.
195
+ # The generated audio is evaluated against target_audio.
196
+ ```
197
+
198
+ ### 6. Attribution and Terms
199
+
200
+ The English and Chinese subsets are derived from CelebV-Dub and CMLR, respectively. Please cite the corresponding datasets and publications when using DubBench.
201
+
202
+ Users must follow the licenses and usage terms of the original datasets and source media. This repository does not grant additional rights to the original audio or video.
203
+
204
+ ## License
205
+
206
+ - **Repository code:** MIT License.
207
+ - **Dataset annotations:** [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
208
+ - **Source media:** Rights remain with the original rights holders; the annotation release does not grant rights to the source videos or audio.