asr.rebble.io: speech recognition for rebble

Rebble ASR provides automatic speech recognition services for Pebble smartwatches.

Uses ElevenLabs' Scribe v1 model for high-quality transcription.

export ASR_API_PROVIDER=elevenlabs
export ASR_API_KEY=your_elevenlabs_api_key

Uses Groq API with Whisper model for fast transcription.

export ASR_API_PROVIDER=groq
export ASR_API_KEY=your_groq_api_key

Uses a local Wyoming-compatible speech recognition service (like Home Assistant's Whisper integration).

export ASR_API_PROVIDER=wyoming-whisper
export WYOMING_HOST=your_wyoming_host  # IP address or hostname
export WYOMING_PORT=10300  # Default Wyoming port

Uses Vosk for offline speech recognition. No API key required.

export ASR_API_PROVIDER=vosk

Enable detailed logging for troubleshooting:

export DEBUG=true

Debug mode provides information about:

• Request details and headers
• Audio processing metrics
• Transcription timing and performance
• Service communication details

Rebble ASR can save audio files and their transcriptions to disk for debugging, quality assurance, or analysis purposes.

export SAVE_RECORDINGS=true
export AUDIO_RECORDINGS_DIR=/path/to/recordings
export MAX_AUDIO_RECORDINGS=10  # Optional, defaults to 10

When enabled, each transcription request will save:

• WAV file: The audio data sent for transcription (e.g., recording_20251002_143022.wav)
• TXT file: The transcribed text (e.g., recording_20251002_143022.txt)

Files are automatically rotated to keep only the most recent recordings based on MAX_AUDIO_RECORDINGS.

When using Docker Compose, map a host directory to persist recordings:

volumes:
  - ./recordings:/recordings

Then set the environment variable to the container path:

export AUDIO_RECORDINGS_DIR=/recordings

• Debugging: Analyze failed or incorrect transcriptions
• Quality Assurance: Review transcription accuracy
• Model Training: Collect real-world audio samples
• Troubleshooting: Identify issues with audio quality or format

• If no API key is provided, falls back to Vosk offline recognition
• If an invalid provider is specified, falls back to Vosk
• If Wyoming-Whisper is selected but the Wyoming package is not installed, falls back to Vosk
• If Wyoming-Whisper fails to connect to the Wyoming service, falls back to Vosk
• Gracefully handles errors by attempting alternative recognition methods