Use Local Whisper

Use Local Whisper

Switches voice transcription from OpenAI's Whisper API to local whisper.cpp. Runs entirely on-device — no API key, no network, no cost.

Channel support: Currently WhatsApp only. The transcription module (src/transcription.ts) uses Baileys types for audio download. Other channels (Telegram, Discord, etc.) would need their own audio-download logic before this skill can serve them.

Note: The Homebrew package is whisper-cpp, but the CLI binary it installs is whisper-cli.

Prerequisites

• voice-transcription skill must be applied first (WhatsApp channel)
• macOS with Apple Silicon (M1+) recommended
• whisper-cpp installed: brew install whisper-cpp (provides the whisper-cli binary)
• ffmpeg installed: brew install ffmpeg
• A GGML model file downloaded to data/models/

Phase 1: Pre-flight

Check if already applied

Read .nanoclaw/state.yaml. If use-local-whisper is in applied_skills, skip to Phase 3 (Verify).

Check dependencies are installed

whisper-cli --help >/dev/null 2>&1 && echo "WHISPER_OK" || echo "WHISPER_MISSING"
ffmpeg -version >/dev/null 2>&1 && echo "FFMPEG_OK" || echo "FFMPEG_MISSING"

If missing, install via Homebrew:

brew install whisper-cpp ffmpeg

Check for model file

ls data/models/ggml-*.bin 2>/dev/null || echo "NO_MODEL"

If no model exists, download the base model (148MB, good balance of speed and accuracy):

mkdir -p data/models
curl -L -o data/models/ggml-base.bin "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin"

For better accuracy at the cost of speed, use ggml-small.bin (466MB) or ggml-medium.bin (1.5GB).

Phase 2: Apply Code Changes

npx tsx scripts/apply-skill.ts .claude/skills/use-local-whisper

This modifies src/transcription.ts to use the whisper-cli binary instead of the OpenAI API.

Validate

npm test
npm run build

Phase 3: Verify

Ensure launchd PATH includes Homebrew

The NanoClaw launchd service runs with a restricted PATH. whisper-cli and ffmpeg are in /opt/homebrew/bin/ (Apple Silicon) or /usr/local/bin/ (Intel), which may not be in the plist's PATH.

Check the current PATH:

grep -A1 'PATH' ~/Library/LaunchAgents/com.nanoclaw.plist

If /opt/homebrew/bin is missing, add it to the <string> value inside the PATH key in the plist. Then reload:

launchctl unload ~/Library/LaunchAgents/com.nanoclaw.plist
launchctl load ~/Library/LaunchAgents/com.nanoclaw.plist

Build and restart

npm run build
launchctl kickstart -k gui/$(id -u)/com.nanoclaw

Test

Send a voice note in any registered group. The agent should receive it as [Voice: <transcript>].

Check logs

tail -f logs/nanoclaw.log | grep -i -E "voice|transcri|whisper"

Look for:

• Transcribed voice message — successful transcription
• whisper.cpp transcription failed — check model path, ffmpeg, or PATH

Configuration

Environment variables (optional, set in .env):

Troubleshooting

"whisper.cpp transcription failed": Ensure both whisper-cli and ffmpeg are in PATH. The launchd service uses a restricted PATH — see Phase 3 above. Test manually:

ffmpeg -f lavfi -i anullsrc=r=16000:cl=mono -t 1 -f wav /tmp/test.wav -y
whisper-cli -m data/models/ggml-base.bin -f /tmp/test.wav --no-timestamps -nt

Transcription works in dev but not as service: The launchd plist PATH likely doesn't include /opt/homebrew/bin. See "Ensure launchd PATH includes Homebrew" in Phase 3.

Slow transcription: The base model processes ~30s of audio in <1s on M1+. If slower, check CPU usage — another process may be competing.

Wrong language: whisper.cpp auto-detects language. To force a language, you can set WHISPER_LANG and modify src/transcription.ts to pass -l $WHISPER_LANG.