@vox-ai-app/vox-voice

ONNX-based wake word detection, keyboard shortcut registration, and voice overlay window for Vox.

Install

npm install @vox-ai-app/vox-voice

Peer dependencies: electron >= 28, @picovoice/pvrecorder-node >= 1.0.0, onnxruntime-node >= 1.18.0

Requirements

• macOS (pvrecorder + ONNX wake word models)
• Microphone permission (requested automatically on first use)
• ONNX models at resources/voice/ in the app package:
  • melspectrogram.onnx
  • embedding_model.onnx
  • computer.onnx

Usage

import {
  initVoiceService,
  destroyVoiceService,
  setLogger,
  setOnWakeWordDetected
} from '@vox-ai-app/vox-voice'

setLogger(logger)
setOnWakeWordDetected(() => {
  // wake word detected — open chat, start listening, etc.
  showChatWindow()
})

await initVoiceService({
  onActivate: () => showChatWindow() // also called on keyboard shortcut
})

// on app quit
await destroyVoiceService()

onActivate handles both wake word detection and the keyboard shortcut (Cmd+Alt+V, falling back to Cmd+Shift+Space). If omitted, only the setOnWakeWordDetected callback fires on wake word.

Voice Window

An always-on-top transparent overlay (400×140px, top-right corner) for displaying voice state.

import { createVoiceWindow, getVoiceWindow, destroyVoiceWindow } from '@vox-ai-app/vox-voice'

createVoiceWindow() // creates the window, loads voice.html
getVoiceWindow() // returns BrowserWindow | null
destroyVoiceWindow() // destroys it

Wake word control

import { pauseWakeWord, resumeWakeWord } from '@vox-ai-app/vox-voice'

pauseWakeWord() // pause detection while user is speaking
resumeWakeWord() // resume after response

Worker entry point

The wake word worker runs in a worker_threads Worker. Register it as a separate entry point in your build config:

// electron.vite.config.js
export default {
  main: {
    build: {
      rollupOptions: {
        input: {
          index: 'src/main/index.js',
          'voice.worker': 'node_modules/@vox-ai-app/vox-voice/src/worker.js'
        }
      }
    }
  }
}

License

MIT