Transcription + slip-of-the-tongue/silence recognition → generate review draft
Quick start
用户: 帮我剪这个口播视频
用户: 处理一下这个视频
Workflow
1. FunASR 30s 分段转录（字符级时间戳）
↓
2. 识别口误（逐句检查）
↓
3. 识别微口误（VAD 检测短片段）
↓
4. 识别语气词（嗯/哎/诶 等）
↓
5. 识别静音（≥1s）
↓
6. 生成审查稿（时间戳驱动）
↓
7. 输出删除任务 TodoList
↓
【等待用户确认】→ 用户确认后，执行 /videocut:剪辑
⚠️ Why use 30s segmentation
FunASR long videos have timestamp drift; 30s segmentation can avoid it.
Progress TodoList
Create at startup:
- [ ] 读取「转录最佳实践」→ 转录视频
- [ ] 读取「口误识别方法论」→ 识别口误
- [ ] VAD 检测微口误（短片段 < 0.5s）
- [ ] 扫描语气词（嗯/哎/诶 等）
- [ ] 识别静音（≥1s）
- [ ] 生成审查稿
- [ ] 输出删除任务清单
⚠️ Must read the methodology before executing
| Stage | Read first | Then execute |
|---|---|---|
| Transcription | tips/转录最佳实践.md |
Call ASR |
| Recognize slips of the tongue | tips/口误识别方法论.md |
Analyze sentence by sentence |
Core: timestamp-driven
Deletion task format
Each itemmust be marked with a precise timestamp (start-end)：
口误（N处）：
- [ ] 1. `(start-end)` 删"错误文本" → 保留"正确文本"
语气词（N处）：
- [ ] 1. `(前字end-后字start)` 删"嗯" 上下文: XX【嗯】YY
静音（N处）：
- [ ] 1. `(start-end)` 静音Xs
Slip-of-the-tongue types
| Type | Example | Deletion strategy |
|---|---|---|
| Repetition type | 拉满新拉满 |
Delete only the difference ("new") |
| Replacement type | AI就是AI就会 |
Delete the first complete version ("AI is") |
| Stutter type | 听会会 |
Delete the first repeated character |
⚠️ Key rules
- Timestamp-driven: The review draft directly marks timestamps, so editing no longer searches the text
- Token-by-token analysis: For slips of the tongue where "delete the earlier part and keep the later part", you must check timestamps token by token
- Check the time span: If the time span of a slip of the tongue is > 2 seconds, there must be silence and it needs to be split
Output files
01-xxx-v1_transcript.json # 转录结果（含时间戳）
01-xxx-v1_审查稿.md # 口误审查稿
Display requirements
After generating the review draft,it must be shown to the user：
- Write to file
01-xxx-v1_审查稿.md - Read and display the content
- Wait for the user to confirm which items to delete
Methodology
See tips/口误识别方法论.md：
- Slip-of-the-tongue recognition method (sentence-by-sentence check)
- Precise handling of "delete the earlier part and keep the later part"
- FunASR timestamp alignment rules