YouTube Transcribe

Smart YouTube video transcription with automatic fallback:

1. Captions first — extracts existing subtitles (manual or auto-generated) via yt-dlp. Fast, free, no compute.
2. Whisper fallback — when no captions exist, downloads audio and transcribes locally with the best available Whisper backend.

When to Use

Use this skill when the user wants to:

• Get a transcript or text version of a YouTube video
• Understand what a YouTube video says without watching it
• Summarize, analyze, or take notes from a YouTube video
• Extract subtitles or captions from a video

Triggers

• "transcribe this YouTube video"
• "what does this video say"
• "get the transcript of [YouTube URL]"
• "summarize this YouTube video" (transcribe first, then process)
• Any YouTube URL shared with a request to understand its content

Requirements

Required:

• yt-dlp — for caption extraction and audio download
• python3

For Whisper fallback (when no captions available):

• ffmpeg — for audio processing
• One of these Whisper backends (auto-detected in priority order):
  1. mlx-whisper — Apple Silicon native, fastest on Mac (pip install mlx-whisper)
  2. faster-whisper — CTranslate2 backend, fast on CUDA/CPU (pip install faster-whisper)
  3. openai-whisper — Original Whisper, universal fallback (pip install openai-whisper)

Usage

Basic — transcribe a video

python3 {baseDir}/scripts/transcribe.py "https://www.youtube.com/watch?v=VIDEO_ID"

Specify language for captions

python3 {baseDir}/scripts/transcribe.py "URL" --language zh

Force Whisper (skip caption check)

python3 {baseDir}/scripts/transcribe.py "URL" --force-whisper

JSON output

python3 {baseDir}/scripts/transcribe.py "URL" --format json

Save to file

python3 {baseDir}/scripts/transcribe.py "URL" --output transcript.txt

Options

Flag	Default	Description
--language	auto	Preferred subtitle/transcription language (e.g. zh, en, ja)
--format	text	Output format: text, json, srt, vtt
--output	stdout	Save transcript to file
--force-whisper	false	Skip caption extraction, go straight to Whisper
--backend	auto	Whisper backend: auto, mlx, faster-whisper, whisper
--model	auto	Whisper model size: auto, large-v3, medium, small, base, tiny

Environment Variables

Variable	Description
YT_WHISPER_BACKEND	Override Whisper backend selection
YT_WHISPER_MODEL	Override Whisper model size

Auto-Detection

Whisper Backend (priority order)

1. MLX Whisper — detected via import mlx_whisper. Best for Apple Silicon.
2. faster-whisper — detected via import faster_whisper. Best for CUDA GPU, good on CPU.
3. OpenAI Whisper — detected via import whisper. Universal fallback.

Model Size (based on available RAM)

RAM	Model	VRAM/RAM Usage
≥16GB	large-v3	~6-10GB
≥8GB	medium	~5GB
≥4GB	small	~2.5GB
<4GB	base	~1.5GB

Caption Language Priority

When --language is not specified, captions are searched in this order:

1. Video's original language
2. Chinese variants: zh-Hant, zh-Hans, zh-TW, zh-CN, zh
3. English: en
4. Any available language

Output Formats

text (default)

Plain text transcript, one continuous block.

json

{
  "video_id": "ZSnYlbIYpjs",
  "title": "Video Title",
  "channel": "Channel Name",
  "duration": 708,
  "language": "zh",
  "method": "captions",
  "transcript": [
    {"start": 0.0, "end": 5.2, "text": "..."},
    ...
  ],
  "full_text": "Complete transcript as single string"
}

srt / vtt

Standard subtitle formats with timestamps.