Video Transcript

Video Transcript Skill

Use this skill when the user asks for a video transcript, video-to-text, YouTube transcript, or a Markdown/text稿 from a video URL or local media file.

Core Policy

1. Do not download full video just to get text. Try caption/subtitle retrieval first.
2. Use captions before ASR. Captions are faster and usually cost nothing.
3. Only download audio for fallback transcription. If captions are missing/unusable, download the best audio stream and run local Whisper.
4. Preserve provenance. Record input URL/file, selected method, language/caption source when known, and output path.
5. For Telegram or chat delivery, prefer a .md attachment for long transcripts. Avoid dumping very long transcripts into one message.

One-command workflow

From this skill directory after yk install video-transcript:

python3 scripts/video_transcript.py "https://www.youtube.com/watch?v=..." \
  --format markdown \
  --output /absolute/path/transcript.md

For a local VTT/SRT caption file:

python3 scripts/video_transcript.py /absolute/path/captions.vtt --format markdown

For a local audio/video file that needs ASR:

python3 scripts/video_transcript.py /absolute/path/talk.mp4 \
  --format markdown \
  --output /absolute/path/talk-transcript.md

Useful options:

• --langs "zh-Hans,zh-CN,en,ja,ko" — caption language priority for yt-dlp.
• --no-asr — captions only; fail instead of downloading audio/transcribing.
• --browser chrome — retry YouTube downloads with --cookies-from-browser chrome when needed.
• --proxy http://127.0.0.1:7890 — pass a proxy to yt-dlp.
• --format markdown|txt|json|srt — choose output shape.
• --caption-file /path/file.vtt — parse a known caption file without network access.
• --asr-engine auto|mlx|faster — local Whisper engine choice.
• --language en|zh|ja|... — language hint for ASR fallback.

Algorithm

input URL/file
  │
  ├─ if local .vtt/.srt or --caption-file
  │    parse caption file → cleaned transcript
  │
  ├─ if URL
  │    yt-dlp captions only
  │      ├─ success: parse VTT/SRT → cleaned transcript
  │      └─ failure/missing captions:
  │            if --no-asr: fail clearly
  │            else download audio only → local Whisper → transcript
  │
  └─ if local media/audio
       local Whisper → transcript

Dependencies

Required for caption parsing only:

• python3 with the standard library.

Required for URL caption/audio retrieval:

• yt-dlp.

Required for ASR fallback:

• One local Whisper engine:
  • Apple Silicon preferred: pip3 install mlx-whisper
  • Portable fallback: pip3 install faster-whisper
• The selected engine may need ffmpeg or its bundled decoder depending on package/version.

The script does not require any API key. API-based ASR can be added later if service-side cost/latency tradeoffs are preferred.

Cost Model

• YouTube/manual/auto captions found: API cost is $0; runtime is usually seconds.
• Local Whisper fallback: API cost is $0; cost is local CPU/GPU time and first model download.
• If later swapped to cloud ASR: current market prices are roughly $0.0025–0.006/minute for common speech-to-text APIs, so a 60-minute video is roughly $0.15–0.36 before any translation/LLM polishing.

For bulk use, cache by video_id + language + output_format + asr_model so repeated requests are free after the first successful transcript.

Verification Checklist

After running:

1. Confirm the command exit code is 0.
2. Confirm stdout or the output file exists and is non-empty.
3. Check the reported Method field:
   • captions / caption-file means no ASR cost.
   • whisper-* means ASR fallback was used.
4. Skim the first and last transcript paragraphs to catch empty/duplicated rolling-caption artifacts.
5. If the output is meant for the user, deliver the .md file or provide a concise excerpt plus the path.

Common Pitfalls

• YouTube can return 403/SABR/PO-token style failures. Retry with --browser chrome and/or --proxy.
• Auto captions may contain rolling/overlapping cues. The script removes adjacent duplicate text and word-overlap when building the plain transcript, but always skim long outputs.
• Do not use full-video download for transcript-only tasks.
• Translation/polishing is a separate task from transcript extraction. Ask before adding translation if the user only requested the original transcript.