zyk21 commited on
Commit
4f2f81b
·
verified ·
1 Parent(s): ae1d22a

Document the DubBench evaluation subset

Browse files
Files changed (1) hide show
  1. README.md +92 -1
README.md CHANGED
@@ -41,8 +41,99 @@ For a configured CUDA environment, replace `--device cpu` with `--device cuda`.
41
 
42
  See the [GitHub README](https://github.com/voxdub-dataset/VoxDub) for full usage details.
43
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
44
  ## License
45
 
46
  - **Repository code:** MIT License.
47
  - **Dataset annotations:** [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
48
- - **Source media:** rights remain with the original rights holders; the annotation release does not grant rights to the source videos or audio.
 
41
 
42
  See the [GitHub README](https://github.com/voxdub-dataset/VoxDub) for full usage details.
43
 
44
+ ## DubBench evaluation subset
45
+
46
+ The `dubbench/` directory is a compact bilingual evaluation collection for systems related to automated dubbing, cross-lingual speech generation, and text-to-speech. Each row pairs a target utterance with a reference utterance and their transcripts. It is intended as evaluation material; it does not define a universal benchmark protocol or leaderboard.
47
+
48
+ The English subset comes from **CelebV-Dub**, introduced with [VoiceCraft-Dub: Automated Video Dubbing with Neural Codec Language Models](https://openaccess.thecvf.com/content/ICCV2025/html/Sung-Bin_VoiceCraft-Dub_Automated_Video_Dubbing_with_Neural_Codec_Language_Models_ICCV_2025_paper.html). See the [official VoiceCraft-Dub project page](https://voicecraft-dub.github.io/) for the paper, examples, code, and data information.
49
+
50
+ The Chinese subset comes from **CMLR (Chinese Mandarin Lip Reading)**. See the [official CMLR dataset page](https://www.vipazoo.cn/CMLR.html) and the associated paper, [A Cascade Sequence-to-Sequence Model for Chinese Mandarin Lip Reading](https://arxiv.org/abs/1908.04917).
51
+
52
+ ### Directory layout
53
+
54
+ ```text
55
+ dubbench/
56
+ |-- english/
57
+ | |-- metadata.tsv # 978 target/reference pairs
58
+ | |-- target/ # 978 target utterances
59
+ | `-- reference/ # 923 unique reference utterances
60
+ `-- chinese/
61
+ |-- metadata.tsv # 1,000 target/reference pairs
62
+ |-- target/ # 1,000 target utterances
63
+ `-- reference/ # 315 unique reference utterances
64
+ ```
65
+
66
+ Utterances normally have three sidecar files with the same stem:
67
+
68
+ - `.wav`: audio waveform
69
+ - `.mp4`: video clip
70
+ - `.txt`: source-side annotation (English files contain a `Text:` record; Chinese files contain the transcript followed by word-level time spans)
71
+
72
+ All target and reference entries have `.wav` and `.txt` files. One English reference, `reference/e71R12fMzfY_5_0`, has no `.mp4` in the supplied source directory; consumers should treat video as optional and check for its presence.
73
+
74
+ ### TSV format
75
+
76
+ The TSV files are UTF-8, have no header row, and contain exactly four tab-separated columns:
77
+
78
+ | Column | Name | Meaning |
79
+ |---:|---|---|
80
+ | 1 | `target_id` | Extensionless target-utterance path relative to the language subset directory, for example `target/-4GNkXEUOwY_6_0_3`. |
81
+ | 2 | `target_text` | Transcript paired with the target utterance. |
82
+ | 3 | `reference_id` | Extensionless paired reference/source-utterance path relative to the language subset directory. |
83
+ | 4 | `reference_text` | Transcript paired with the reference/source utterance. |
84
+
85
+ There are no reserved, empty, or audio-root columns in the supplied files. The language subset directory is the path root. Append `.wav`, `.mp4`, or `.txt` to an ID to locate the corresponding asset.
86
+
87
+ Example row (spaces are shown around the tab separators for readability):
88
+
89
+ ```text
90
+ target/-4GNkXEUOwY_6_0_3 <TAB> THERE'S NOTHING TO INSTALL, PATCH, OR UPGRADE EVER <TAB> reference/-4GNkXEUOwY_0_1_0 <TAB> AND WHILE IT IS EASY ...
91
+ ```
92
+
93
+ ### Download and parse DubBench
94
+
95
+ Download only the DubBench directory with the Hugging Face CLI:
96
+
97
+ ```bash
98
+ hf download zyk21/VoxDub --repo-type dataset --include "dubbench/**" --local-dir VoxDub
99
+ ```
100
+
101
+ Or download it programmatically and resolve each extensionless ID against its language directory:
102
+
103
+ ```python
104
+ import csv
105
+ from pathlib import Path
106
+ from huggingface_hub import snapshot_download
107
+
108
+ repo_root = Path(snapshot_download(
109
+ repo_id="zyk21/VoxDub",
110
+ repo_type="dataset",
111
+ allow_patterns="dubbench/**",
112
+ ))
113
+ subset_root = repo_root / "dubbench" / "english" # or "chinese"
114
+
115
+ with (subset_root / "metadata.tsv").open(encoding="utf-8", newline="") as f:
116
+ for target_id, target_text, reference_id, reference_text in csv.reader(f, delimiter="\t"):
117
+ target_wav = subset_root / f"{target_id}.wav"
118
+ target_video = subset_root / f"{target_id}.mp4"
119
+ reference_wav = subset_root / f"{reference_id}.wav"
120
+
121
+ assert target_wav.is_file()
122
+ assert reference_wav.is_file()
123
+ if target_video.is_file():
124
+ pass # Load or evaluate the target video here.
125
+ ```
126
+
127
+ Use the IDs in the TSV rather than enumerating directory order: reference utterances may be reused by multiple rows.
128
+
129
+ ### DubBench attribution and source terms
130
+
131
+ DubBench reorganizes evaluation pairs whose English material comes from CelebV-Dub and whose Chinese material comes from CMLR. Please cite the corresponding source work when using a subset.
132
+
133
+ The source datasets may impose research-only, redistribution, or other use restrictions. This repository does not replace or broaden the licenses or terms of CelebV-Dub, its underlying sources, or CMLR, and it grants no additional rights to source media. Users are responsible for reviewing and complying with the current terms published by the original dataset owners before downloading, redistributing, or using the files.
134
+
135
  ## License
136
 
137
  - **Repository code:** MIT License.
138
  - **Dataset annotations:** [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
139
+ - **Source media:** rights remain with the original rights holders; the annotation release does not grant rights to the source videos or audio.