Skip to main content
A realtime voice provider plugin handles audio I/O: capturing audio, transcribing speech to text, and synthesizing text to speech.

​
Registration

import { createPlugin } from "openclaw/plugin-sdk"

export default createPlugin({
  name: "my-voice-provider",
  config: { /* config schema */ },
  setup(sdk, config) {
    sdk.registerRealtimeVoiceProvider({
      name: "my-voice-provider",

      async connect({ userId, channelId }) {
        return {
          sessionId: crypto.randomUUID(),
          async sendAudio(chunk: Uint8Array) { /* send to model */ },
          async close() { /* tear down */ },
          onTranscript(cb) { /* call cb with transcribed segments */ },
          onAudio(cb) { /* call cb with synthesized chunks */ },
        }
      },

      async disconnect(session) {
        await session.close()
      },
    })
  },
})

​
VoiceSession interface

MethodDescription
sendAudio(chunk: Uint8Array)Send a raw PCM audio chunk (16-bit, 16kHz, mono)
onTranscript(cb)Register callback for transcription results
onAudio(cb)Register callback to receive synthesized audio
close()Tear down the session and release resources

​
Audio format

Input and output: 16-bit PCM, 16 kHz sample rate, mono. Convert other formats before passing to sendAudio.

​
Error handling

Throw from connect() to signal that the provider cannot accept a session. The gateway falls back to text-only mode.