phattbeats Meeting Transcription Skill
Transcribes meeting audio files using faster-whisper, generates executive summary, and saves to vault.
Installation
faster-whisper is pre-installed at:
/root/.openclaw/faster-whisper-env/
To reinstall:
curl -LsSf https://astral.sh/uv/install.sh | sh
source $HOME/.local/bin/env
uv venv /root/.openclaw/faster-whisper-env
/root/.openclaw/faster-whisper-env/bin/pip install faster-whisper
Scripts
scripts/transcribe.py— Transcribe a single audio filescripts/meeting-pipeline.py— Full pipeline: transcribe → summarize → save to vault
Usage
Transcribe an audio file
/root/.openclaw/faster-whisper-env/bin/python /root/.openclaw/workspace/skills/meeting-transcription/scripts/transcribe.py <audio_file> [options]
Options:
--model tiny|base|small|medium|large-v3— Model size (default: base)--language LANG— Language code (default: en)--device cpu|cuda— Device (default: cpu)--output FILE— Output JSON file
Full pipeline
export LITELLM_API_KEY=sk-phatt-litellm-salt-permanent-2026
/root/.openclaw/faster-whisper-env/bin/python \
/root/.openclaw/workspace/skills/meeting-transcription/scripts/meeting-pipeline.py \
<audio_file> "Client Name" "Meeting Title" \
--date 2026-03-29
Output:
- Prints vault path on success
- Returns JSON with transcript, summary, action items
Audio File Formats
Supported: WAV, MP3, M4A, OGG, FLAC (ffmpeg converts automatically if available). Without ffmpeg: WAV and MP3 work directly.
Model Selection
| Model | Size | Speed | Accuracy |
|---|---|---|---|
| tiny | ~75MB | Fastest | Baseline |
| base | ~150MB | Fast | Good |
| small | ~500MB | Medium | Better |
| medium | ~1.5GB | Slow | High |
| large-v3 | ~3GB | Slowest | Best |
Default: base — Good balance of speed and accuracy for most meetings.
Testing
Test with real audio:
/root/.openclaw/faster-whisper-env/bin/python \
/root/.openclaw/workspace/skills/meeting-transcription/scripts/transcribe.py \
/path/to/real/meeting.m4a \
--model small \
--output /tmp/transcript.json
Troubleshooting
"No speech detected":
- Audio may be too quiet, too short, or in a different language
- Try
--language enexplicitly - Try a larger model:
--model medium
Model download slow:
- Models cached at
~/.cache/huggingface/ - ~150MB for base model, first run downloads
CUDA/GPU:
- If GPU available:
--device cuda - RTX 4090 or better recommended for real-time transcription