| name | whatsapp-voice-transcriber |
| description | Use when a user sends a voice note and wants it transcribed, or says "transcribe this voice note", "what did I say?", "convert voice to text", or "transcribe this audio". Handles OGG, MP3, WAV, and other audio formats. Outputs transcription with content classification (Idea/Task, Draft, or Note). Uses mlx_whisper large-v3-turbo for Apple Silicon — fully local, no API cost. |
| metadata | {"emoji":"🎙️","category":"automation","platform":"macOS (Apple Silicon)"} |
WhatsApp Voice Transcriber
Transcribes voice notes locally using mlx_whisper (Apple Silicon) and classifies by content type. No API cost, runs fully offline.
Requirements
- mlx_whisper (recommended):
pip install mlx-whisper — uses large-v3-turbo model on Apple Silicon
- Fallback: openai-whisper (
pip install openai-whisper) — base model
- Python 3.10+
- OpenClaw CLI (optional, for send-back delivery)
Usage
python3 scripts/transcribe.py <audio_file_path>
Example:
python3 scripts/transcribe.py /tmp/voice_note.ogg
Configuration (env vars)
| Variable | Default | Description |
|---|
MLX_WHISPER_PATH | /opt/homebrew/bin/mlx_whisper | Path to mlx_whisper binary |
WHISPER_MODEL | mlx-community/whisper-large-v3-turbo | Whisper model to use |
NOTIFY_TARGET | (empty) | Phone/JID to send result to via openclaw |
NOTIFY_CHANNEL | whatsapp | Channel for notification |
OPENCLAW_PATH | /opt/homebrew/bin/openclaw | Path to openclaw binary |
Set NOTIFY_TARGET to auto-send transcription after processing:
NOTIFY_TARGET="+1234567890" python3 scripts/transcribe.py /tmp/voice.ogg
Classification
Heuristic classifier labels each transcription:
- 💡 Idea / Task — keywords: remind, task, build, idea, need to
- ✉️ Draft — keywords: tell, message, send, write, reply
- 📝 Note — everything else
OpenClaw Integration
Wire into your agent skill dispatcher to auto-transcribe incoming voice notes:
import subprocess, tempfile, os
def handle_voice_note(attachment_path):
result = subprocess.run(
["python3", "~/clawd/.agents/skills/whatsapp-voice-transcriber/scripts/transcribe.py",
attachment_path],
capture_output=True, text=True
)
return result.stdout.strip()
Model Notes
- mlx_whisper large-v3-turbo — best for Apple Silicon (M1/M2/M3). Highly accurate, fast enough for real-time. ~1GB model download on first use.
- whisper base — fallback only. Much lower accuracy. Use if no Apple Silicon.
- Model is downloaded automatically on first run.
Scripts
scripts/transcribe.py — Main script. Transcribes audio, classifies, prints, and optionally sends.
