Video No Subtitle Transcribe

无字幕视频转写兜底方案。当视频（YouTube/Bilibili 等）没有字幕、字幕接口被禁或拉取失败时，用 yt-dlp 下载音频 + faster-whisper 本地转写，输出带时间戳的完整文稿。触发词：视频没字幕、转写视频、字幕被禁、whisper 转写、提取语音内容。

When subtitle-based approaches (e.g. bilibili-youtube-watcher) fail — video has no CC/auto-generated subtitles, or the subtitle endpoint is disabled by YouTube ("Subtitles are disabled for this video") — use this skill: download audio → local speech recognition → timestamped transcript.

Use Cases

• YouTube/Bilibili videos without subtitles (creator disabled CC or auto-generated captions)
• Subtitle endpoints rate-limited/disabled by YouTube
• Any yt-dlp-supported site, or local audio/video files

Prerequisites

• yt-dlp (pip install yt-dlp or brew install yt-dlp)
• faster-whisper (pip install faster-whisper)
• ffmpeg (system package)
• whisper small model: ~/.local/share/whisper-small/model.bin (auto-downloaded from ModelScope if missing, ~483MB)

Usage

bash

python3 {baseDir}/scripts/transcribe_video.py "<URL>" --lang zh --out /tmp/transcript.txt

Arguments:

• --lang: language code, zh for Chinese, en for English (default zh)
• --out: output file path (default transcript.txt)
• --proxy: proxy URL, e.g. http://127.0.0.1:7890 (default: env vars HTTPS_PROXY/HTTP_PROXY, then OpenClaw config)

Output format: one line per segment, [start-end seconds] text, e.g. [2.0-4.0] Hello everyone, welcome to today's session

Workflow (full pipeline)

1. User provides a video link → first try the subtitle path (bilibili-youtube-watcher)
2. Subtitle path fails (no subs / disabled / fetch error) → use this skill
3. Download audio with this script (yt-dlp client fallbacks run automatically).
4. If every yt-dlp download attempt fails, use the browser fallback below to obtain an audio/video file, then run this script again with that local file path.
5. After transcription, read the output transcript → understand content by timeline → structured summary for the user

Browser download fallback (only after yt-dlp fails)

Use this only when the script reports [dl] FAILED; do not send a video link to a third-party downloader before trying the local yt-dlp path.

1. Open https://youtube.iiilab.com/ in the browser.
2. Paste the YouTube video URL into the site, start its parse/download flow, and select the best available audio format (preferred) or a video file.
3. Download the resulting file to a local path. If the site requests login, a CAPTCHA, payment, permissions, or any acknowledgement that data will be sent to a third party, stop and ask the user before proceeding.
4. Transcribe the downloaded local file; no yt-dlp download is needed:

bash

python3 {baseDir}/scripts/transcribe_video.py "<LOCAL_AUDIO_OR_VIDEO_FILE>" --lang zh --out /tmp/transcript.txt

Privacy note: This fallback submits the source URL to youtube.iiilab.com, a third-party service. Tell the user before using it for private, unlisted, sensitive, or client-owned videos. Do not use it for non-YouTube URLs; report the yt-dlp failure instead.

Pitfalls (must follow)

1. YouTube's default client gets 429/bot-checked ("Sign in to confirm you're not a bot") → try audio clients in order: tv_embedded (most reliable, bypasses DRM/SABR experiments, verified) → android → ios → default. The Android client may hit the SABR-only experiment (formats without URLs) — normal, just try the next client.

2. Use the ModelScope CN mirror for model download (~1MB/s): https://modelscope.cn/models/Systran/faster-whisper-small/resolve/master/model.bin. Don't use HuggingFace direct (very slow from China via proxy, ~90KB/s), and don't use the hf-mirror endpoint (its xet protocol silently corrupts files).

3. Verify model integrity: the intact file is exactly 483546902 bytes. If it doesn't match, delete and re-download; if whisper errors with File model.bin is incomplete: failed to read a buffer... the file is corrupt. Don't use a fixed size as the "done" signal — let the request finish naturally and verify the byte count.

4. Proxy: YouTube requires a proxy (mainland China); ModelScope direct access needs no proxy. Proxy priority: --proxy argument > env vars > OpenClaw config.

5. Transcription params: WhisperModel(dir, device='cpu', compute_type='int8') + transcribe(audio, language=lang, vad_filter=True). Speed is roughly 0.5–1x realtime (1-hour audio takes ~30–60 min). Be patient, don't kill the process early.

6. Homophone typos (e.g. 在/再, 做/作 confusion) are normal for whisper small — restore by context when summarizing.

Limitations

• Transcription quality depends on audio clarity; fast speech/noise cause errors
• Slower than cloud transcription services (minutes vs seconds), but completely free, runs locally, no privacy leakage

License

MIT