AIVoices_HA

Custom Piper text-to-speech voices for Home Assistant.

These are medium-quality, single-speaker English (en-US) neural voices trained for use with Home Assistant's local TTS engine.

Voices

Voice	Files	Language	Sample rate	Quality
cathy	en_us-cathy-medium.onnx + .onnx.json	en-US	22050 Hz	medium
cynda	en_us-cynda-medium.onnx + .onnx.json	en-US	22050 Hz	medium
james	en_us-james-medium.onnx + .onnx.json	en-US	22050 Hz	medium

Each voice consists of two files that must stay together:

• *.onnx — the trained model weights
• *.onnx.json — the voice configuration (phoneme map, sample rate, inference settings)

More voices coming

This is an ongoing collection — additional Piper voices will be added over time. Star or watch the repository to be notified as new voices land.

Installation (Home Assistant)

These voices work with the Piper add-on / wyoming-piper.

1. Locate the Piper data directory where custom voices are stored. For the official Piper add-on this is the share folder, typically:

   /share/piper/
   

2. Copy both files for each voice you want into that directory:

   en_us-cathy-medium.onnx
   en_us-cathy-medium.onnx.json
   

3. Restart the Piper add-on. Then in Home Assistant go to Settings → Voice assistants → (your assistant) → Text-to-speech and select the new voice.

Tip: If a voice does not appear, confirm the .onnx and .onnx.json filenames match exactly and that both files are present.

Testing a voice locally

With Piper installed you can render a sample to a WAV file:

echo "Hello from Home Assistant." | \
  piper --model en_us-cathy-medium.onnx --output_file sample.wav

AI usage disclaimer

Claude (by Anthropic) is used in this repository for repository management and documentation — tasks such as setting up the git repo, writing and maintaining this README, and organizing files. The Piper voice models themselves are not generated by Claude.

License

The licensing of a trained Piper voice depends on the dataset it was trained on. Add a LICENSE file describing the terms that apply to these models and their training data before distributing them.