Event Bus

The normalized event bus is Kith's primary contract. Adapters emit these events; consumers subscribe to them. Everything else is an implementation detail.

Event Types

Event	When	Key Fields
turn_start	A new conversational turn begins	turnId, role ("user" or "assistant")
turn_end	A turn completes	turnId, role
tts_start	TTS synthesis begins for a chunk	turnId, chunkId
tts_audio_chunk	A slice of synthesized audio	turnId, chunkId, audioB64, mimeType
tts_end	TTS synthesis completes for a chunk	turnId, chunkId
stt_partial	Partial speech-to-text transcription	turnId, text
stt_final	Final speech-to-text transcription	turnId, text
viseme_frame	A single viseme for lip sync	turnId, viseme, weight, offsetMs
emotion_state	Emotional state hint for avatar	state, intensity
barge_in_detected	User interrupted the assistant	turnId
reconnect	Runtime reconnected after disconnect	attempt
error	Something went wrong	message, retriable

Subscribing

const unsub = voice.on((event) => {
  switch (event.type) {
    case "tts_audio_chunk":
      // Play audio
      break;
    case "emotion_state":
      // Tint avatar
      break;
    case "barge_in_detected":
      // Stop playback
      break;
  }
});

// Later: unsub();

Event Flow

A typical voice.speak("Hello! haha") produces this sequence:

emotion_state { state: "neutral", intensity: 0.5 }
turn_start    { turnId: "t1", role: "assistant" }
tts_start     { turnId: "t1", chunkId: "c1" }
tts_audio_chunk × N
tts_end       { turnId: "t1", chunkId: "c1" }
turn_end      { turnId: "t1", role: "assistant" }

All events include a timestamp field (Unix ms).

Router vs Runtime Events

• Runtime events come from the sidecar: turn_start, turn_end, tts_*, stt_*, viseme_frame, barge_in_detected, reconnect, error.
• Router-synthesized events come from VoiceRouter: emotion_state (parsed from emojis in the input text).

Subscribe via voice.on() (not runtime.on()) to get both.