| --- |
| title: Speech-free & Music-free Clips — Random Sample |
| emoji: 🎲 |
| colorFrom: blue |
| colorTo: green |
| sdk: static |
| app_file: index.html |
| pinned: false |
| license: mit |
| --- |
| |
| # Speech-free & music-free source clips — a random sample of 12 |
|
|
| Drawn at random (seed 20260809) from the **8,191** clips that survived a two-stage audio filter over |
| **101,503** source videos from InsAVE-80K and JAVEdit-100k: |
|
|
| 1. **No speech** — FireRedVAD over every clip, then FireRedASR2-AED transcription on every VAD |
| positive; a clip is kept only if no real words were transcribed. |
| 2. **No music** — no music-related AudioSet tag (74 of them: Music, Background music, Musical |
| instrument, Theme music, Guitar, Piano…) anywhere in the top five. |
|
|
| These are the **inputs** to the object-removal pipeline, not its outputs. |
|
|
| ## Sampled, not cherry-picked |
|
|
| 3 of the 12 carry `Speech` as their top AudioSet tag despite VAD and ASR both clearing them. That is |
| the known model disagreement: FireRedASR2 covers Chinese and English, so speech in another language |
| transcribes empty and slips through. Those are flagged in the manifest as `speech_disagreement`; the |
| stricter variant (`kept_no_speech_strict.csv`, 6,983 clips) drops them. |
|
|
| Manifest: `/group2/ct/weihanx/JavisDiT/javisdit/script/extra_train.json` |
|
|
