dvcrn Douyin Video Transcribe
Extract speech from Douyin videos and convert to text. Supports Chinese/English, cross-platform (Windows/macOS/Linux).
Core Principle
Douyin has strict anti-scraping. Must:
- Load page in browser, wait for video stream
- Extract real CDN URL from DOM or network requests
- Download with
Referer: https://www.douyin.com/header (403 without it) - Convert audio to 16kHz mono WAV for Whisper
Prerequisites
| Tool | Purpose | Install |
|---|---|---|
| ffmpeg | Audio extraction | brew install ffmpeg / winget install ffmpeg / apt install ffmpeg |
| whisper | Speech-to-text | pip install openai-whisper |
| curl | Download video | Built-in (Windows: curl.exe) |
Workflow
1. Resolve Short URL
Douyin share links are usually v.douyin.com/xxx, resolve to full URL:
# macOS/Linux
curl -sL -o /dev/null -w '%{url_effective}' "https://v.douyin.com/xxx/"
# Windows PowerShell
curl.exe -sL -o NUL -w "%{url_effective}" "https://v.douyin.com/xxx/"
Output: https://www.douyin.com/video/7616020798351871284
Video ID is the 19-digit number in URL.
2. Get Video URL
Open video page in browser, wait 3-5 seconds, execute JS:
(() => {
const videos = document.querySelectorAll('video');
for (const v of videos) {
const src = v.currentSrc || v.src;
if (src && src.startsWith('http') && !src.includes('uuu_265')) {
return src;
}
}
return null;
})()
Key points:
- Returns
null: Page not loaded, retry after waiting - Contains
uuu_265: Placeholder video, retry after waiting - Starts with
blob:: Streaming, wait for real URL - CDN URLs expire (~2 hours), re-fetch if needed
3. Download Video
# macOS/Linux
curl -L -H "Referer: https://www.douyin.com/" -o video.mp4 "<CDN_URL>"
# Windows
curl.exe -L -H "Referer: https://www.douyin.com/" -o video.mp4 "<CDN_URL>"
Referer header is required, otherwise 403.
4. Extract Audio
ffmpeg -i video.mp4 -ar 16000 -ac 1 -c:a pcm_s16le audio.wav -y
Parameters:
-ar 16000: 16kHz sample rate (Whisper requirement)-ac 1: Mono channel-c:a pcm_s16le: 16-bit PCM
5. Transcribe
python -m whisper audio.wav --model small --language zh
Model selection:
| Model | Size | 5-min video (CPU) | Accuracy | Use case |
|---|---|---|---|---|
| tiny | 75MB | ~30s | Fair | Quick preview |
| base | 142MB | ~1min | Good | Daily use |
| small | 466MB | ~3min | Better | Recommended |
| medium | 1.5GB | ~8min | Best | High accuracy |
Language:
- Chinese:
--language zh - English:
--language en - Auto-detect: omit flag (slower)
Output files in current directory: audio.txt, audio.srt, audio.json
Troubleshooting
| Issue | Detection | Solution |
|---|---|---|
| Short URL fails | Returns non-douyin.com | Check link completeness, remove share text noise |
| Video URL not found | JS returns null | Wait 3-5s and retry, max 3 times |
| Placeholder video | URL contains uuu_265 |
Page not loaded, wait and retry |
| Download 403 | curl returns 403 | Check Referer header; URL may be expired |
| Whisper hangs | No output for long time | First run downloads model (~460MB for small) |
| Garbled output | Terminal shows gibberish | Normal, read .txt file directly |
| Out of memory | Process killed | Use smaller model (base/tiny) |
Output Convention
Name files by video ID, save to user-specified directory:
output/
├── 7616020798351871284.mp4 # Original video (optional)
├── 7616020798351871284.wav # Audio (delete after)
├── 7616020798351871284.txt # Transcript
└── 7616020798351871284.srt # Subtitles (optional)
Scripts (Optional)
Helper scripts in skill directory:
scripts/get_video_url.js: Browser-side video URL extraction with multiple methodsscripts/transcribe.py: CLI one-click transcription (requires video URL)
Scripts are accelerators, not required. Implement yourself after understanding the workflow.
Notes
- Article links (/article/): Use browser snapshot directly, no transcription needed
- Douyin AI summary: Some video pages have AI-generated chapter summaries, extract from snapshot as supplement
- Other platforms: This skill is for Douyin only. Use yt-dlp for YouTube/Bilibili