video-understanding

Video Understanding Skill

Download videos and transcribe their content for analysis.

Prerequisites

• yt-dlp installed (pip install yt-dlp or brew install yt-dlp)
• ffmpeg installed (brew install ffmpeg or apt install ffmpeg)
• Whisper installed (pip install openai-whisper)

Pipeline

1. Download Video

# Download video with yt-dlp
yt-dlp -o "assets/downloads/%(title)s.%(ext)s" "<VIDEO_URL>"

For best quality

yt-dlp -f "bestvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best"
-o "assets/downloads/%(title)s.%(ext)s" "<VIDEO_URL>"

For audio only (faster)

yt-dlp -x --audio-format mp3
-o "assets/downloads/%(title)s.%(ext)s" "<VIDEO_URL>"

### 2. Extract Audio (if downloaded video)

```bash
ffmpeg -i "assets/downloads/video.mp4" \
  -vn -acodec mp3 -ab 128k \
  "assets/downloads/audio.mp3"

3. Transcribe with Whisper

# Basic transcription
whisper "assets/downloads/audio.mp3" \
  --model base \
  --output_format txt \
  --output_dir output/

# Higher quality (slower)
whisper "assets/downloads/audio.mp3" \
  --model medium \
  --output_format all \
  --output_dir output/

# With timestamps
whisper "assets/downloads/audio.mp3" \
  --model base \
  --output_format srt \
  --output_dir output/

4. Read and Analyze Transcript

Read the generated transcript file from output/
Summarize key points
Extract quotes and timestamps
Identify speakers if multiple

Model Options

Model	Size	Speed	Quality
tiny	39M	Fastest	Lower
base	74M	Fast	Good
small	244M	Medium	Better
medium	769M	Slow	High
large	1550M	Slowest	Highest

Output Formats

• txt - Plain text transcript
• srt - SubRip subtitles with timestamps
• vtt - WebVTT subtitles
• json - Detailed JSON with word-level timing
• all - All formats

Tips

• Use base model for speed, medium for accuracy
• Add --language en to force English detection
• Use --task translate to translate to English
• Check assets/downloads/ for downloaded files
• Store transcripts in output/transcripts/