voice

Overview ¶

Package voice provides transport-agnostic voice capabilities for AI calls.

It is deliberately independent of WebSocket, WebRTC, SIP, or any other media transport. Transports feed PCM (or encoded audio decoded upstream) into a voice session and receive synthesized audio back through callbacks.

Layering:

protocol/voice/dialog    session orchestration + Event/Command contract
protocol/voice/gateway   dialog-plane WebSocket client
protocol/voice/webrtc    browser WebRTC (HTTP SDP + SRTP/Opus)
protocol/voice/xiaozhi   xiaozhi-esp32 / browser WebSocket (pipeline + realtime)
protocol/voice/transport per-call SessionFactory wiring
protocol/voice/asr       uplink pipeline components
protocol/voice/tts       downlink synthesis + cache

Uplink order matters: VAD runs on raw microphone PCM before echo suppression so barge-in still works while the recognizer feed is silenced. An optional Denoiser (RNNoise, WebRTC AEC3, hardware AEC) may run after decode and before VAD. PlaybackGate tracks streaming, queued TTS, and a post-playback tail for room-echo suppression when true AEC is unavailable.

Conversation logic (LLM, tools, business rules) lives outside this package. The dialog subpackage defines the event/command contract between the voice plane and an external Dialog application.

Typical integration:

sess, _ := dialog.NewSession(ctx, dialog.Config{
    CallID: "call-1",
    Engine: recognizerEngine,
    TTSService: tts.FromSynthesisEngine(synth),
    OnAudioOut: transport.SendDownlink,
    OnEvent:    dialogApp.HandleEvent,
})
sess.Start(ctx)
transport.OnUplink(func(pcm []byte) { sess.ProcessAudio(ctx, pcm) })
dialogApp.OnCommand(sess.HandleCommand)