mofa-youtube

name	mofa-youtube
version	0.1.0
description	Download YouTube video or transcribe local audio files via ASR. Triggers: mofa youtube, transcribe video, youtube to text, 转录视频, YouTube转文字, extract transcript, video to text, 提取字幕, download youtube, 下载视频, transcribe, 转录, transcribe audio, transcribe mp3, 音频转文字, speech to text, ASR, 语音识别.
requires_bins	yt-dlp,ffmpeg,ffprobe,curl,jq
requires_env	null

mofa-youtube

Extracts structured content from a YouTube video: transcript, key frames, and metadata. Zero Python — all shell + curl + ffmpeg.

Output Paths

Always use a unique per-request directory:

skill-output/mofa-youtube-<YYYYMMDD-HHMMSS>/
├── metadata.json            # {title, url, duration, chapters, topics, language}
├── transcript.txt           # Full ASR transcript
├── transcript_chapters/     # Per-chapter transcripts (if video has chapters)
│   ├── ch01.txt
│   └── ch02.txt
├── images/
│   ├── frame-01.png
│   ├── frame-02.png
│   ├── frame-03.png
│   └── frame-04.png
└── video.mp4                # Original (deleted unless keep_video=true)

Onboarding

Required:

• yt-dlp — YouTube download. Install: brew install yt-dlp or pip install yt-dlp
• ffmpeg / ffprobe — audio extraction, frame capture, chunking. Install: brew install ffmpeg
• curl + jq — ASR API calls
• Local ASR server at localhost:8080 running qwen3-asr model

Check before running:

yt-dlp --version
ffmpeg -version
curl -s http://localhost:8080/v1/models | jq .

Pipeline

download → extract_audio → transcribe ──┐
    └──→ extract_frames ────────────────┼──→ deliver
                                summarize┘

Step	What it does	Model
download	yt-dlp downloads video + metadata JSON	cheap
extract_audio	ffmpeg converts to 16kHz mono WAV	cheap
transcribe	Chunks audio (4-min), base64 + POST to ASR	cheap
extract_frames	Smart frame extraction: chapter timestamps (preferred) or scene change detection (fallback)	cheap
summarize	LLM reads transcript, generates metadata.json	strong
deliver	Sends transcript, frames, and summary back to chat via message + send_file	cheap

ASR Details

• Endpoint: http://localhost:8080/v1/audio/transcriptions
• Model: qwen3-asr
• Chunk size: 240 seconds (4 minutes)
• Audio format: 16kHz, mono, PCM s16le WAV
• Request: JSON with base64-encoded file, model, and language fields
• Timeout: 300 seconds per chunk

Transcription shell commands

IMPORTANT: Each 4-min audio chunk base64-encodes to ~10MB, which exceeds shell argument limits. You MUST use file-based curl -d @file, never pass the base64 inline.

# Get duration
dur=$(ffprobe -v quiet -show_entries format=duration -of csv=p=0 audio.wav | cut -d. -f1)

# Chunk and transcribe
for start in $(seq 0 240 "$dur"); do
  ffmpeg -y -i audio.wav -ss "$start" -t 240 \
    -acodec pcm_s16le -ar 16000 -ac 1 /tmp/asr_chunk.wav 2>/dev/null
  # Base64 encode to file (strip newlines)
  base64 < /tmp/asr_chunk.wav | tr -d '\n' > /tmp/asr_chunk.b64
  # Build JSON by concatenating around the base64 file
  { printf '{"file":"'; cat /tmp/asr_chunk.b64; printf '","model":"qwen3-asr","language":"en"}'; } > /tmp/asr_req.json
  # POST using -d @file to avoid arg limits
  curl -s -X POST http://localhost:8080/v1/audio/transcriptions \
    -H 'Content-Type: application/json' --max-time 300 \
    -d @/tmp/asr_req.json \
    | jq -r '.text' >> transcript.txt
done

Composability

This skill outputs a content package that mofa-site consumes directly:

mofa-youtube → mofa-site → mofa-publish

Also useful standalone for: transcription, subtitle generation, key frame extraction for mofa-slides.

metadata.json Schema

{
  "title": "The Essence of Calculus, Chapter 1",
  "url": "https://youtube.com/watch?v=WUvTyaaNkzM",
  "duration_secs": 1087,
  "language": "en",
  "topics": ["derivatives", "area under curve", "fundamental theorem"],
  "chapters": [
    {"title": "Introduction", "start_secs": 0},
    {"title": "The circle problem", "start_secs": 45}
  ],
  "key_concepts": ["limit", "infinitesimal", "integral"],
  "frame_timestamps": [163, 413, 674, 924]
}

Timing

Video Length	Chunks	Estimated Time
5 min	2	~1 min
20 min	5	~3 min
60 min	15	~8 min

Bottleneck is ASR (sequential chunk processing). Download and frame extraction are fast.