web/hype.js

// hype.js — the VoxCPM2 AI hype-man. A voice bank designed in a club-MC persona
// (Spanglish energy, on-brand and Latin-flavored) is triggered INSTANTLY by
// deterministic musical events and ducked over the mix. The model owns the
// VOICE; deterministic code owns WHEN it speaks — the "her" pattern.
//
// Graceful fallback so the show always goes on:
//   1. pre-baked VoxCPM2 WAVs at  hype/<key>.wav   (bulletproof, instant)
//   2. browser SpeechSynthesis                      (last resort, dev/offline)

export const HYPE_LINES = {
  hello: "Manos arriba — let's go!",
  build: "Build it up… here it comes…",
  drop:  "¡Dale! Drop it!",
  sweep: "Bring it up!",
  mix:   "Mix it — vamos!",
};

function b64ToBuf(b64) {
  const bin = atob(b64), u = new Uint8Array(bin.length);
  for (let i = 0; i < bin.length; i++) u[i] = bin.charCodeAt(i);
  return u.buffer;
}

export class HypeMan {
  constructor(engine) {
    this.engine = engine;
    this.ctx = engine.ctx;
    this.bank = {};                 // key -> AudioBuffer (VoxCPM2)
    this.last = {};                 // key -> last fire time (ctx seconds)
    this.enabled = true;
    // usable immediately via browser speech; upgrades to VoxCPM2 as load() fills the bank
    this.source = ("speechSynthesis" in window) ? "speech" : "none";
    this.bus = this.ctx.createGain();
    this.bus.gain.value = 1.0;
    this.bus.connect(this.ctx.destination);
  }

  // Non-blocking: prefer pre-baked static WAVs, then the live ZeroGPU endpoint.
  // Any key not yet loaded simply falls back to browser speech in fire().
  async load() {
    let loaded = await this._loadPrebaked();
    if (loaded < Object.keys(HYPE_LINES).length) loaded += await this._loadEndpoint();
    if (loaded) this.source = "voxcpm";
    return this.source;
  }

  async _loadPrebaked() {
    let n = 0;
    for (const key of Object.keys(HYPE_LINES)) {
      try {
        const r = await fetch(`hype/${key}.wav`, { cache: "force-cache" });
        if (r.ok) { this.bank[key] = await this.ctx.decodeAudioData(await r.arrayBuffer()); n++; }
      } catch (_) { /* not deployed — fall through */ }
    }
    return n;
  }

  async _loadEndpoint() {
    try {
      const r = await fetch("api/hype");        // VoxCPM2 on ZeroGPU (cached server-side)
      if (!r.ok) return 0;
      const j = await r.json();
      let n = 0;
      for (const item of (j.bank || [])) {
        if (this.bank[item.key]) continue;
        try { this.bank[item.key] = await this.ctx.decodeAudioData(b64ToBuf(item.wav_b64)); n++; } catch (_) {}
      }
      return n;
    } catch (_) { return 0; }
  }

  fire(key, { cooldown = 5, duck = 0.4 } = {}) {
    if (!this.enabled || !HYPE_LINES[key]) return;
    const now = this.ctx.currentTime;
    if (this.last[key] && now - this.last[key] < cooldown) return;
    this.last[key] = now;

    if (duck < 1) this.engine.duck(duck, 1.1);

    const buf = this.bank[key];
    if (buf) {
      const s = this.ctx.createBufferSource();
      s.buffer = buf; s.connect(this.bus); s.start();
    } else if (this.source === "speech") {
      const u = new SpeechSynthesisUtterance(HYPE_LINES[key]);
      u.rate = 1.08; u.pitch = 1.15; u.volume = 1.0;
      window.speechSynthesis.cancel();
      window.speechSynthesis.speak(u);
    }
  }
}