asr-timing-review

name	asr-timing-review
description	Use for global recorded-video transcript and timing work: run ASR, inspect timestamped transcript segments, infer intended speech, manually decide visual cue timestamps, and produce timing-cues.json without relying on automatic cue matching alone.

ASR Timing Review

ASR is evidence for editorial timing, not the final authority.

Required Loop

1. Run the local ASR command when available: LESSON_ID=<lesson-id> npm run asr:recording
2. Read shared/composer/<lesson-id>/asr-transcript.json.
3. Compare the raw ASR with the intended script, recording context, and audio/video timing.
4. Write a human-review ledger:
   • time range
   • raw ASR text
   • inferred intended speech
   • visual event
   • confidence
   • review note
5. Manually write timing-cues.json from the reviewed ledger.
6. Update render-plan shot starts/durations only after the ledger explains why each cue time is chosen.

Rules

• NEVER treat programmatic cue matching as final timing when ASR is weak.
• NEVER match only sparse English tech tokens in Chinese recordings.
• Use rough script timestamps only as prior expectations, not final timing.
• If the ASR is crowded but the idea is inferable, state the inferred intended speech and use the timestamp range.
• If the idea is not inferable, mark needs_manual_review and do not call the render final.
• Prefer cue anchors that the speaker can say clearly next time, such as 第一条：动画时钟, 第二条：浏览器资产, 警告：不是迁移口号, and 最后：声音驱动画布.

Output Shape

Create or update:

• shared/composer/<lesson-id>/asr-human-review.md
• shared/composer/<lesson-id>/timing-cues.json
• render plan timings that reference timingSource: human-reviewed ASR ledger