Voice widgets

import { createBarVisualizer } from '/widgets/shared/audio-viz.js';

const agentId = new URL(location.href).searchParams.get('agentId');
await host.ready();

const attach = await host.live.attach(`voice:${agentId}`, {
  transports: ['portal-relay'],
});
const t = attach.transport; // { kind:'portal-relay', relayUrl, relayToken }

const ws = new WebSocket(`${t.relayUrl}&token=${encodeURIComponent(t.relayToken)}`);

ws.onmessage = (ev) => {
  const msg = JSON.parse(ev.data);
  if (msg.type === 'session.created') { /* ready to talk */ }
  if (msg.type === 'response.audio.delta') playWavChunk(msg.audio); // base64 WAV, 24kHz
  if (msg.type === 'response.done') { /* agent finished its turn */ }
  if (msg.type === 'response.cancelled') stopPlayback(); // barge-in acknowledged
};

// Push-to-talk: stream PCM16 mono 16kHz while held, then commit.
function onHold() { /* mic → ws.send({ type:'input.audio', audio: base64Pcm16 }) */ }
function onRelease() { ws.send(JSON.stringify({ type: 'input.audio_committed' })); }

// Barge-in: interrupt the agent mid-response.
function interrupt() { ws.send(JSON.stringify({ type: 'response.cancel' })); }

await host.ready();
const attach = await host.live.attach(`voice:${agentId}`);
// attach.meta.iframeBootstrap.models === { llm, voiceId, systemPrompt }
const { llm, voiceId, systemPrompt } = attach.meta.iframeBootstrap.models;

// Dynamic import + WebGPU pipeline setup; no portal proxying.
const { pipeline } = await import('https://cdn.jsdelivr.net/npm/@huggingface/transformers');
const stt = await pipeline('automatic-speech-recognition', 'onnx-community/whisper-small');
const tts = await pipeline('text-to-speech', `onnx-community/Kokoro-82M-${voiceId}`);