react-native-voice-activator

Appearance

Examples ​

Two different shapes, for the two ways this package is used.

Voice agent handoffs (mic gate) ​

The package listens on-device, and hands a live microphone to a voice agent that does everything else server-side. No sttProvider, no ttsProvider: those vendors already do both halves, so the managed session would be redundant.

Both turn on the same rule: the microphone has one owner at a time. Stop detection before the vendor claims the audio session, and start it again on every exit path, including the failures. An agent session that never connects otherwise leaves the app deaf with no visible symptom.

Provider adapters (managed session) ​

These demonstrate how to keep STT and TTS integrations outside the package core while still using the public provider interfaces:

  • SpeechToTextProvider
  • TextToSpeechProvider
  • initialize({ sttProvider, ttsProvider, autoSpeak })

Architecture rule: reference adapters live in docs/examples/ and app-level code only. User-owned custom adapters belong in application code, not in the library package. They do not belong in src/, src/internal/, or the mandatory package runtime.

Model training (custom wake word) ​

To train a Sherpa-ONNX KWS bundle and keyword file for engineConfig.assetKeys, see ../model-training/wake-word-training.md. Training is offline; the package loads ONNX + keywords.txt only.

Model training (custom TTS voice) ​

To train or fine-tune a Piper voice and export .onnx for CustomTTSConfig.modelPath, see ../model-training/tts-voice-cloning.md. Training is offline; the app supplies a local model file and a matching phonemize implementation.

Current Reference Adapters ​

What These Examples Prove ​

  • provider vendors remain application-owned
  • the package public API stays engine-neutral and vendor-neutral by default
  • consumers can compose wake word detection with STT/TTS without reaching into private runtime internals
  1. validate the package-owned wake-word runtime first
  2. layer in an application-owned STT adapter for the transcribe step
  3. layer in an application-owned TTS adapter for the optional speak step

The example app shows the real wake-word runtime, links back to these adapter docs, and includes separate simulated host-provider previews for the optional transcribe and speak steps.

It also exposes a bundled keyword selector that applies preset engineConfig.assetKeys.keywordAssetKey values for the shipped Sherpa phrases, rather than generating or editing custom keyword files at runtime.

What These Examples Do Not Prove ​

  • this repo does not automate real vendor audio capture, recognition, or speech playback
  • Expo Go is still unsupported for this package runtime
  • vendor install steps, permissions, and compatibility constraints must still be validated in the host app that adopts these adapters