Custom TTS Provider (CustomTTSAdapter)

CustomTTSAdapter is the built-in on-device TTS adapter that runs Piper TTS ONNX models through onnxruntime-react-native. It implements TextToSpeechProvider and can be passed directly to initialize({ ttsProvider }).

Do not move it into the library package core. Phonemize logic is model-specific and belongs in your application.

Install

npm install onnxruntime-react-native

Implement a phonemize callback

Piper TTS ONNX models require espeak-ng phoneme IDs as BigInt64Array. You must supply a phonemize callback that converts input text to the correct IDs for your model. Options:

• Use piper-phonemize (a JS/WASM port of espeak-ng) if available for your target platform.
• Implement a lookup table for a fixed vocabulary.
• Call a server-side phonemizer if on-device phonemization is not required.

const phonemize = async (text: string): Promise<BigInt64Array> => {
  // Example: call your phonemizer here
  const ids = await myPhonemizer.textToIds(text);
  return new BigInt64Array(ids);
};

Wire the adapter

import {
  initialize,
  startDetection,
  CustomTTSAdapter,
  WhisperRNSTTAdapter,
} from 'react-native-voice-activator';

const ttsProvider = new CustomTTSAdapter({
  modelPath: '/path/to/voice.onnx', // absolute local path after model download
  phonemize,
  sampleRate: 22050, // default — omit for standard Piper models
  // speakerId: 0,   // only needed for multi-speaker Piper models
});

await initialize({
  sttProvider: new WhisperRNSTTAdapter({ modelId: 'whisper-tiny-en' }),
  ttsProvider,
  autoSpeak: true,
});
await startDetection();

Model sourcing

CustomTTSAdapter does not download models. You must bundle or download the .onnx model file yourself and pass its absolute local path as modelPath.

• Download Piper voice models from the official Piper releases page.
• For voice cloning, follow docs/model-training/tts-voice-cloning.md (Story 13-2).
• Store the model in the app's document directory or bundle it as a native asset.

Stop / barge-in

adapter.stop() cancels any in-progress synthesis. The adapter sets a cancellation flag that is checked at every async checkpoint inside speak(), so the native audio ring buffer is flushed promptly.