companion/@applications/web/src/worklets/mic-processor.js

/**
 * MicProcessor AudioWorkletProcessor
 *
 * Captures microphone audio at the browser's native sample rate,
 * resamples to 16kHz mono via linear interpolation, and emits
 * 960-sample Int16 frames (30ms at 16kHz) to the main thread.
 *
 * Wire format (ArrayBuffer):
 *   [0x01][seq: 4B big-endian][pcm: 960 * 2 bytes = 1920 bytes Int16]
 * Total frame size: 1925 bytes
 */

const TARGET_SAMPLE_RATE = 16000;
const FRAME_SAMPLES = 960; // 30ms at 16kHz
const HEADER_BYTES = 5;    // 1 type byte + 4 seq bytes
const FRAME_BYTES = FRAME_SAMPLES * 2; // Int16 = 2 bytes per sample
const BUFFER_BYTES = HEADER_BYTES + FRAME_BYTES;

class MicProcessor extends AudioWorkletProcessor {
  constructor() {
    super();

    /** Resampled PCM accumulation buffer */
    this._accumulator = new Float32Array(FRAME_SAMPLES * 2);
    /** Write position in accumulator */
    this._accPos = 0;
    /** Fractional sample position for linear interpolation resampling */
    this._resamplePhase = 0.0;
    /** Per-frame sequence counter (uint32, wraps) */
    this._seq = 0;
    /** Ratio: input samples per output sample */
    this._ratio = 0;
  }

  /**
   * Linear interpolation resampler.
   * Takes a block of float32 input samples (at native rate) and appends
   * resampled float32 output samples (at 16kHz) into the accumulator,
   * flushing complete 960-sample frames to the main thread.
   *
   * @param {Float32Array} input - Input samples at native sample rate
   */
  _resampleAndAccumulate(input) {
    const inputLen = input.length;
    if (inputLen === 0) return;

    // Compute ratio on first real input block (sampleRate is available in worklet)
    if (this._ratio === 0) {
      this._ratio = sampleRate / TARGET_SAMPLE_RATE;
    }

    const ratio = this._ratio;
    let phase = this._resamplePhase;

    while (phase < inputLen) {
      const i0 = Math.floor(phase);
      const i1 = Math.min(i0 + 1, inputLen - 1);
      const frac = phase - i0;

      const s0 = input[i0];
      const s1 = input[i1];
      const sample = s0 + frac * (s1 - s0);

      this._accumulator[this._accPos++] = sample;

      if (this._accPos === FRAME_SAMPLES) {
        this._flushFrame();
        this._accPos = 0;
      }

      phase += ratio;
    }

    // Carry over fractional phase (subtract consumed integer samples)
    this._resamplePhase = phase - inputLen;
  }

  /**
   * Encode and post a complete 960-sample frame.
   */
  _flushFrame() {
    const buffer = new ArrayBuffer(BUFFER_BYTES);
    const view = new DataView(buffer);

    // Header
    view.setUint8(0, 0x01);
    view.setUint32(1, this._seq, false); // big-endian

    // PCM: Float32 → Int16 with clamping.
    // Int16Array requires 2-byte-aligned offsets; HEADER_BYTES=5 is odd, so use a
    // separate aligned buffer and copy the bytes in with Uint8Array.
    const pcmTemp = new Int16Array(FRAME_SAMPLES);
    for (let i = 0; i < FRAME_SAMPLES; i++) {
      const f = this._accumulator[i];
      const clamped = f < -1.0 ? -1.0 : f > 1.0 ? 1.0 : f;
      pcmTemp[i] = Math.round(clamped * 32767);
    }
    new Uint8Array(buffer, HEADER_BYTES).set(new Uint8Array(pcmTemp.buffer));

    this.port.postMessage(buffer, [buffer]);
    this._seq = (this._seq + 1) >>> 0; // unsigned 32-bit increment
  }

  /**
   * @param {Float32Array[][]} inputs
   * @returns {boolean}
   */
  process(inputs) {
    const input = inputs[0];
    if (!input || input.length === 0) return true;

    // Mix down to mono if stereo
    const ch0 = input[0];
    if (!ch0 || ch0.length === 0) return true;

    let mono;
    if (input.length > 1 && input[1] && input[1].length === ch0.length) {
      mono = new Float32Array(ch0.length);
      const ch1 = input[1];
      for (let i = 0; i < ch0.length; i++) {
        mono[i] = (ch0[i] + ch1[i]) * 0.5;
      }
    } else {
      mono = ch0;
    }

    this._resampleAndAccumulate(mono);
    return true;
  }
}

registerProcessor('mic-processor', MicProcessor);