audio-worklet.js

// audio-worklet.js — streaming PCM player for the kitsunechat TTS bridge.
//
// Input: chunks of Float32 mono PCM at the AudioContext's native rate
// (already resampled by tts-bridge.js). We queue them and write them to
// the output buffer in order, sample-by-sample.
//
// NO fade-in/out. Pocket TTS emits many small chunks (one per diffusion
// generation_step). Applying a per-chunk fade creates audible dips to
// silence at every chunk boundary — the "air gaps every 1 second" bug.
// Match the tts-site worklet: straight FIFO copy, gapless by construction.
//
// 'clear' message: drop the queue + reset state (used when active message
// changes; stops any stale OLD-gen audio from playing).
// 'finish' message: set `finishing = true`; once queue drains, post 'ended'.

class StreamingAudioProcessor extends AudioWorkletProcessor {
  constructor() {
    super();
    this.queue = [];        // Float32Array chunks (FIFO, already at native rate)
    this.offset = 0;        // read position in head chunk
    this.finishing = false;
    this.framesProcessed = 0;
    this.lastRmsReport = 0;
    this.port.onmessage = (e) => {
      const msg = e.data;
      if (msg.type === 'chunk') {
        this.queue.push(msg.samples);
      } else if (msg.type === 'finish') {
        this.finishing = true;
      } else if (msg.type === 'clear') {
        this.queue = [];
        this.offset = 0;
        this.finishing = false;
      }
    };
  }

  process(inputs, outputs) {
    const out = outputs[0][0];
    if (!out || out.length === 0) return true;

    let written = 0;
    while (written < out.length && this.queue.length > 0) {
      const chunk = this.queue[0];
      const available = chunk.length - this.offset;
      const needed = out.length - written;
      const n = Math.min(available, needed);

      // Straight copy — no envelope. Adjacent chunks play continuously.
      out.set(chunk.subarray(this.offset, this.offset + n), written);

      written += n;
      this.offset += n;
      if (this.offset >= chunk.length) {
        this.queue.shift();
        this.offset = 0;
      }
    }

    for (let i = written; i < out.length; i++) out[i] = 0;

    // Diagnostics: post RMS level of the output buffer every ~250ms so the
    // main thread can confirm audio is being produced (and not silent).
    this.framesProcessed++;
    if (this.framesProcessed - this.lastRmsReport >= 12) {
      this.lastRmsReport = this.framesProcessed;
      let sumSq = 0;
      for (let i = 0; i < out.length; i++) sumSq += out[i] * out[i];
      const rms = Math.sqrt(sumSq / out.length);
      const peak = 0;
      this.port.postMessage({ type: 'level', rms, peak, frame: this.framesProcessed });
    }

    if (this.finishing && this.queue.length === 0) {
      this.port.postMessage({ type: 'ended' });
      this.finishing = false;
    }

    return true;
  }
}

registerProcessor('streaming-audio', StreamingAudioProcessor);