Subtitles
Transcription → Agent Proofreading → Manual Review → Burning
Core Process (approximately 8-15 minutes total, including manual review)
1. Extract audio + upload ⏱ ~1min
↓
2. Volcengine transcription (with hotwords) ⏱ ~2min
↓
3. Agent automatic proofreading ⏱ ~3-5min
↓
4. Manual review confirmation ⏱ depends on user
↓
5. Burn subtitles ⏱ ~1-2min
Step 1: Extract audio and upload
# Extract audio
ffmpeg -i "video.mp4" -vn -acodec libmp3lame -y audio.mp3
# Upload to uguu.se (temporary file hosting)
curl -s -F "files[][email protected]" https://uguu.se/upload
# Returns URL like: https://o.uguu.se/xxxxx.mp3
Step 2: Volcengine transcription (with hotwords)
The transcription script automatically reads the dictionary as hotwords to improve recognition accuracy:
# Dictionary location: /Users/chengfeng/Desktop/AIos/剪辑Agent/.claude/skills/字幕/词典.txt
# The script loads it automatically
bash ../剪口播/scripts/volcengine_transcribe.sh "https://o.uguu.se/xxxxx.mp3"
Dictionary format (one word per line):
skills
Claude
Agent
成峰
剪辑skills
claude code
Step 3: Agent automatic proofreading
3.1 Generate subtitles with timestamps
const result = JSON.parse(fs.readFileSync('volcengine_result.json'));
const subtitles = result.utterances.map((u, i) => ({
id: i + 1,
text: u.text,
start: u.start_time / 1000,
end: u.end_time / 1000
}));
fs.writeFileSync('subtitles_with_time.json', JSON.stringify(subtitles, null, 2));
3.2 Manual proofreading (no scripts)
After transcription, the Agent must read through all subtitles line by line and manually correct the following issues:
Common misrecognition rules table
| Misrecognition | Correct | Type |
|---|---|---|
| 成风 | 成峰 | Homophone |
| 正特/整特 | Agent | Misrecognition |
| IT就 | Agent就 | Similar pronunciation |
| edge的叉100 | Agentx100 | Misrecognition |
| cloud code | Claude Code | Similar pronunciation |
| Schill/skill | skills | Similar pronunciation |
| 剪口拨/剪口波 | 剪口播 | Homophone |
| 自净化/资金化 | 自进化 | Homophone |
| 减口播 | 剪口播 | Homophone |
| 录剪 | 漏剪 | Homophone |
| 作为这个 | 做这个 | Homophone |
| 斜杠V1/斜杠v | /v | Spoken description |
| 斜杠v点口拨 | /v.口播 | Spoken + homophone |
| a p i t/APIK | API Key | Misrecognition |
| excuse | skills | Misrecognition |
| 移完了 | 剪完了 | Homophone |
Common missing word issues
| Original | Corrected | Description |
|---|---|---|
| 步呢是配置 | 第二步呢是配置 | Missing "第二" |
| 4步就是 | 第4步就是 | Missing "第" |
| 别省时间 | 特别省时间 | Missing "特" |
| 这个我们的 | 这个是我们的 | Missing "是" |
| 在里面看到 | 可以在里面看到 | Missing "可以" |
| 跟大家处理完 | 跟大家讲完 | Wrong word choice |
| 剪辑了逻辑 | 剪辑的逻辑 | Grammar error |
3.3 Compare with original script (if available)
If there is an original script/manuscript, it can assist proofreading, but do not use scripts for automatic matching (text differences can cause cumulative timestamp errors).
The Agent should:
- Read the original script as a reference
- Compare each line manually, correcting differences as they are found
- Mark uncertain spots for manual review
Step 4: Start review server
cd subtitle_directory/
node /path/to/skills/字幕/scripts/subtitle_server.js 8898 "video.mp4"
Access http://localhost:8898
Features:
- Left side: video playback, right side: subtitle list
- Auto-highlights current subtitle during playback
- Double-click subtitle text to edit (timestamps remain unchanged)
- Variable speed playback (1x/1.5x/2x/3x)
- Save subtitles / Export SRT / Burn subtitles
- Bottom bar shows dictionary quick insert
Step 5: Burn subtitles
Default style: 22pt golden bold, black outline 2px, centered at bottom
ffmpeg -i "video.mp4" \
-vf "subtitles='video.srt':force_style='FontSize=22,FontName=PingFang SC,Bold=1,PrimaryColour=&H0000deff,OutlineColour=&H00000000,Outline=2,Alignment=2,MarginV=30'" \
-c:a copy \
-y "video_subtitled.mp4"
| Parameter | Value | Description |
|---|---|---|
| FontSize | 22 | Font size |
| FontName | PingFang SC | PingFang font |
| Bold | 1 | Bold |
| PrimaryColour | &H0000deff | Golden #ffde00 |
| OutlineColour | &H00000000 | Black outline |
| Outline | 2 | Outline width |
| Alignment | 2 | Centered at bottom |
| MarginV | 30 | Bottom margin |
Directory structure
output/YYYY-MM-DD_videoname/subtitles/
├── 1_transcription/
│ ├── audio.mp3
│ └── volcengine_result.json
├── subtitles_with_time.json # Core file
└── 3_output/
├── video.srt
└── video_subtitled.mp4
Subtitle rules
| Rule | Description |
|---|---|
| One line per screen | No line breaks, no stacking |
| No punctuation at end of sentence | Hello not Hello. |
| Keep punctuation within sentence | Click here, then there |
Feedback log
2026-01-31
- Volcengine supports hotwords, already integrated into transcription script
- Agent must proofread automatically after transcription; cannot hand over directly to user
- Subtitle style: golden bold #ffde00, outline 2px
- "IT" is often misrecognized as "Agent", needs error correction rule added
- Important: Agent proofreading must be done manually line by line; cannot use scripts for automatic matching
- Added 17 common misrecognition rules (see section 3.2)
- Missing words are harder to detect than misrecognitions; special attention required