Spaces:

townbox
/

ava

Running

ava / workers /vad-segmenter.js

Remove repeat guard, restore conservative speech detection, drop persona name line

8545065 verified 4 days ago

2.8 kB

// @ts-check
/**
 * Turns per-frame speech probabilities into utterances. Pure logic (no model,
 * no audio I/O) so it can be unit tested.
 *
 * Two thresholds: normally a short confident run opens an utterance, but while
 * she is talking ("guarded") the bar is higher and longer, so a cough or leaked
 * speaker audio doesn't cut her off while a real interruption still does.
 */

export const FRAME_SAMPLES = 512; // 32 ms at 16 kHz

export const DEFAULTS = {
  startProb: 0.5,
  startFrames: 4, // 128 ms
  guardedStartProb: 0.75,
  guardedStartFrames: 9, // 288 ms
  endProb: 0.35,
  endFrames: 17, // 544 ms of quiet ends the turn
  trailFrames: 6, // keep 192 ms of the quiet tail
  prerollFrames: 12, // 384 ms before the trigger, so onsets aren't clipped
  minSpeechFrames: 8, // 256 ms of actual speech, or it was a misfire
  maxFrames: 938, // 30 s
};

export class Segmenter {
  /** @param {Partial<typeof DEFAULTS>} [opt] */
  constructor(opt = {}) {
    this.opt = { ...DEFAULTS, ...opt };
    this.guarded = false;
    this.reset();
  }

  reset() {
    /** @type {Float32Array[]} */
    this._preroll = [];
    /** @type {Float32Array[]} */
    this._speech = [];
    this._inSpeech = false;
    this._run = 0;
    this._quiet = 0;
    this._voiced = 0;
  }

  /**
   * @param {Float32Array} frame
   * @param {number} prob
   * @returns {{ kind: "start" } | { kind: "end", samples: Float32Array } | { kind: "misfire" } | null}
   */
  push(frame, prob) {
    const o = this.opt;
    if (!this._inSpeech) {
      this._preroll.push(frame);
      if (this._preroll.length > o.prerollFrames + o.guardedStartFrames) this._preroll.shift();
      const startProb = this.guarded ? o.guardedStartProb : o.startProb;
      const startFrames = this.guarded ? o.guardedStartFrames : o.startFrames;
      this._run = prob >= startProb ? this._run + 1 : 0;
      if (this._run < startFrames) return null;
      this._inSpeech = true;
      this._speech = this._preroll;
      this._preroll = [];
      this._voiced = this._run;
      this._quiet = 0;
      this._run = 0;
      return { kind: "start" };
    }

    this._speech.push(frame);
    if (prob >= o.endProb) {
      this._quiet = 0;
      this._voiced += 1;
    } else {
      this._quiet += 1;
    }
    if (this._quiet >= o.endFrames || this._speech.length >= o.maxFrames) {
      const dropTail = Math.max(0, this._quiet - o.trailFrames);
      const kept = this._speech.slice(0, this._speech.length - dropTail);
      const voiced = this._voiced;
      this.reset();
      if (voiced < o.minSpeechFrames) return { kind: "misfire" };
      const samples = new Float32Array(kept.length * FRAME_SAMPLES);
      kept.forEach((f, i) => samples.set(f, i * FRAME_SAMPLES));
      return { kind: "end", samples };
    }
    return null;
  }
}

	// @ts-check
	/**
	* Turns per-frame speech probabilities into utterances. Pure logic (no model,
	* no audio I/O) so it can be unit tested.
	*
	* Two thresholds: normally a short confident run opens an utterance, but while
	* she is talking ("guarded") the bar is higher and longer, so a cough or leaked
	* speaker audio doesn't cut her off while a real interruption still does.
	*/

	export const FRAME_SAMPLES = 512; // 32 ms at 16 kHz

	export const DEFAULTS = {
	startProb: 0.5,
	startFrames: 4, // 128 ms
	guardedStartProb: 0.75,
	guardedStartFrames: 9, // 288 ms
	endProb: 0.35,
	endFrames: 17, // 544 ms of quiet ends the turn
	trailFrames: 6, // keep 192 ms of the quiet tail
	prerollFrames: 12, // 384 ms before the trigger, so onsets aren't clipped
	minSpeechFrames: 8, // 256 ms of actual speech, or it was a misfire
	maxFrames: 938, // 30 s
	};

	export class Segmenter {
	/** @param {Partial<typeof DEFAULTS>} [opt] */
	constructor(opt = {}) {
	this.opt = { ...DEFAULTS, ...opt };
	this.guarded = false;
	this.reset();
	}

	reset() {
	/** @type {Float32Array[]} */
	this._preroll = [];
	/** @type {Float32Array[]} */
	this._speech = [];
	this._inSpeech = false;
	this._run = 0;
	this._quiet = 0;
	this._voiced = 0;
	}

	/**
	* @param {Float32Array} frame
	* @param {number} prob
	* @returns {{ kind: "start" } \| { kind: "end", samples: Float32Array } \| { kind: "misfire" } \| null}
	*/
	push(frame, prob) {
	const o = this.opt;
	if (!this._inSpeech) {
	this._preroll.push(frame);
	if (this._preroll.length > o.prerollFrames + o.guardedStartFrames) this._preroll.shift();
	const startProb = this.guarded ? o.guardedStartProb : o.startProb;
	const startFrames = this.guarded ? o.guardedStartFrames : o.startFrames;
	this._run = prob >= startProb ? this._run + 1 : 0;
	if (this._run < startFrames) return null;
	this._inSpeech = true;
	this._speech = this._preroll;
	this._preroll = [];
	this._voiced = this._run;
	this._quiet = 0;
	this._run = 0;
	return { kind: "start" };
	}

	this._speech.push(frame);
	if (prob >= o.endProb) {
	this._quiet = 0;
	this._voiced += 1;
	} else {
	this._quiet += 1;
	}
	if (this._quiet >= o.endFrames \|\| this._speech.length >= o.maxFrames) {
	const dropTail = Math.max(0, this._quiet - o.trailFrames);
	const kept = this._speech.slice(0, this._speech.length - dropTail);
	const voiced = this._voiced;
	this.reset();
	if (voiced < o.minSpeechFrames) return { kind: "misfire" };
	const samples = new Float32Array(kept.length * FRAME_SAMPLES);
	kept.forEach((f, i) => samples.set(f, i * FRAME_SAMPLES));
	return { kind: "end", samples };
	}
	return null;
	}
	}