homeassistant-mcp/extra

Speech-to-Text Examples

This directory contains examples demonstrating how to use the speech-to-text integration with wake word detection.

Prerequisites

1. Make sure you have Docker installed and running
2. Build and start the services:

   docker-compose up -d
   

Running the Example

1. Install dependencies:

   npm install
   

2. Run the example:

   npm run example:speech
   

   Or using ts-node directly:

   npx ts-node examples/speech-to-text-example.ts
   

Features Demonstrated

1. Wake Word Detection

   • Listens for wake words: "hey jarvis", "ok google", "alexa"
   • Automatically saves audio when wake word is detected
   • Transcribes the detected speech

2. Manual Transcription

   • Example of how to transcribe audio files manually
   • Supports different models and configurations

3. Event Handling

   • Wake word detection events
   • Transcription results
   • Progress updates
   • Error handling

Example Output

When a wake word is detected, you'll see output like this:

🎤 Wake word detected!
  Timestamp: 20240203_123456
  Audio file: /path/to/audio/wake_word_20240203_123456.wav
  Metadata file: /path/to/audio/wake_word_20240203_123456.wav.json

📝 Transcription result:
  Full text: This is what was said after the wake word.

  Segments:
    1. [0.00s - 1.52s] (95.5% confidence)
       "This is what was said"
    2. [1.52s - 2.34s] (98.2% confidence)
       "after the wake word."

Customization

You can customize the behavior by:

1. Changing the wake word models in docker/speech/Dockerfile
2. Modifying transcription options in the example file
3. Adding your own event handlers
4. Implementing different audio processing logic

Troubleshooting

1. Docker Issues

   • Make sure Docker is running
   • Check container logs: docker-compose logs fast-whisper
   • Verify container is up: docker ps

2. Audio Issues

   • Check audio device permissions
   • Verify audio file format (WAV files recommended)
   • Check audio file permissions

3. Performance Issues

   • Try using a smaller model (tiny.en or base.en)
   • Adjust beam size and patience parameters
   • Consider using GPU acceleration if available