RuLips-1k: quality report
Computed from manifest.jsonl (423,047 clips, 1134.4 h).
| part | RuTube clips | RuTube h | MultiVSR clips | MultiVSR h |
|---|---|---|---|---|
| train | 68,079 | 409.7 | 343,149 | 693.5 |
| val | 1,705 | 6.4 | 3,035 | 6.0 |
| test | 953 | 6.1 | 6,126 | 12.6 |
Clip length
- rutube: median 9.0 s, p10 2.6, p90 63.1, max 119.9
- multivsr: median 7.5 s, p10 2.5, p90 11.7, max 14.0
Face and head pose (from landmarks)
- rutube: face found in 99.6% of frames on average; clips with <85% coverage: 0.8%; |azimuth| > 30°: 16.0%, > 60°: 0.17%; |zenith| > 30°: 0.2%; mouth opening median 0.056
- multivsr: face found in 94.9% of frames on average; clips with <85% coverage: 8.2%; |azimuth| > 30°: 8.6%, > 60°: 0.60%; |zenith| > 30°: 0.4%; mouth opening median 0.065
Audio-visual sync and transcript (RuTube)
- SyncNet confidence: median 4.33, p10 3.11 (clips below 2.75 were dropped at collection)
- ASR disagreement (CER between window-level and per-clip transcripts): median 0.072; ≥ 0.5 in 8.9% of clips
- Burned-in subtitles: 11.3% of clips have subtitles in more than half of the frames
- Words per clip: median 20; per-clip transcripts: 100.0%
Speakers
- rutube: 933 speakers; top-10 share 9.2%; median speaker 13.6 min
- multivsr: 7,197 speakers; top-10 share 0.7%; median speaker 4.9 min