@tensamin/audio 0.1.13 → 0.1.15

package/README.md

@@ -101,11 +101,9 @@ vad: {
   energyVad?: {
     smoothing: number;                 // Default: 0.95
     initialNoiseFloor: number;         // Default: 0.001
-    noiseFloorAdaptRateQuiet: number;  // Default: 0.01
-    noiseFloorAdaptRateLoud: number;   // Default: 0.1
-    minSNR: number;                    // Default: 10.0 (dB)
-    snrRange: number;                  // Default: 10.0 (dB)
-    minEnergy: number;                 // Default: 0.001
+    minSNR: number;                    // Default: 8.0 (dB)
+    snrRange: number;                  // Default: 12.0 (dB)
+    minEnergy: number;                 // Default: 0.01
   };
 }
 ```
@@ -116,7 +114,7 @@ vad: {
 - `stopThreshold`: Probability threshold to mute audio (Default: 0.3, ~13dB SNR)
 - `hangoverMs`: Delay before muting after speech stops (Default: 300ms)
 - `preRollMs`: Audio buffer duration before speech onset
-- `minSpeechDurationMs`: Minimum duration to consider as valid speech (Default: 150ms)
+- `minSpeechDurationMs`: Minimum duration to consider as valid speech (Default: 250ms)
 - `minSilenceDurationMs`: Minimum silence duration between speech segments
 
 **Energy VAD Parameters:**
@@ -124,7 +122,7 @@ vad: {
 - `smoothing`: Energy calculation smoothing factor (0-1)
 - `minSNR`: Minimum signal-to-noise ratio in dB for speech detection
 - `snrRange`: Range in dB for probability scaling from minSNR
-- `minEnergy`: Minimum absolute RMS energy to consider as speech (Default: 0.001, ~-60dB)
+- `minEnergy`: Minimum absolute RMS energy to consider as speech (Default: 0.01, ~-40dB)
 
 ### Output Control
 

package/dist/{chunk-FKR6NWZF.mjs → chunk-GLKAWCEW.mjs}

@@ -3,11 +3,9 @@ var createEnergyVadWorkletCode = (vadConfig) => {
   const energyParams = vadConfig?.energyVad || {};
   const smoothing = energyParams.smoothing ?? 0.95;
   const initialNoiseFloor = energyParams.initialNoiseFloor ?? 1e-3;
-  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.01;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.1;
-  const minSNR = energyParams.minSNR ?? 10;
-  const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 1e-3;
+  const minSNR = energyParams.minSNR ?? 8;
+  const snrRange = energyParams.snrRange ?? 12;
+  const minEnergy = energyParams.minEnergy ?? 0.01;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -15,8 +13,6 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     this.smoothing = ${smoothing};
     this.energy = 0;
     this.noiseFloor = ${initialNoiseFloor};
-    this.noiseFloorAdaptRateQuiet = ${noiseFloorAdaptRateQuiet};
-    this.noiseFloorAdaptRateLoud = ${noiseFloorAdaptRateLoud};
     this.minSNR = ${minSNR};
     this.snrRange = ${snrRange};
     this.minEnergy = ${minEnergy};
@@ -36,8 +32,11 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     
     // Calculate instantaneous RMS (Root Mean Square) energy
     let sum = 0;
+    let peak = 0;
     for (let i = 0; i < channel.length; i++) {
+      const sample = Math.abs(channel[i]);
       sum += channel[i] * channel[i];
+      peak = Math.max(peak, sample);
     }
     const instantRms = Math.sqrt(sum / channel.length);
 
@@ -45,32 +44,21 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     // this.energy acts as the smoothed RMS value
     this.energy = this.energy * this.smoothing + instantRms * (1 - this.smoothing);
 
-    // Adaptive noise floor estimation
-    // We use a TWO-PASS approach to avoid circular dependencies:
-    // FIRST PASS: Calculate instantaneous SNR to decide how to adapt
-    const instantSnr = instantRms / (this.noiseFloor + 1e-6);
-    const instantSnrDb = 20 * Math.log10(Math.max(1e-6, instantSnr));
-    
-    // Adapt the noise floor based on instantaneous SNR
-    if (instantRms < this.noiseFloor) {
-      // Signal is quieter than noise floor, adapt downwards quickly
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
-    } else if (instantSnrDb < 12) {
-      // Signal is louder but SNR is low (< 12dB) - likely just louder background noise
-      // Adapt upwards at normal rate to track rising noise
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateLoud) + instantRms * this.noiseFloorAdaptRateLoud;
-    } else {
-      // Signal has high SNR (>= 12dB) - likely speech or transient
-      // Adapt VERY slowly to avoid "chasing" speech
-      const slowRate = this.noiseFloorAdaptRateLoud * 0.02;
-      this.noiseFloor = this.noiseFloor * (1 - slowRate) + instantRms * slowRate;
-    }
+    // Calculate Crest Factor (peak-to-RMS ratio)
+    // Voice typically has crest factor of 2-4 (6-12dB)
+    // Keyboard clicks have crest factor of 10-30+ (20-30dB)
+    const crestFactor = peak / (instantRms + 1e-10);
+    const crestFactorDb = 20 * Math.log10(Math.max(1e-6, crestFactor));
 
-    // Ensure noise floor doesn't drop to absolute zero
-    // 0.00005 is approx -86dB, very quiet but prevents SNR explosion
-    this.noiseFloor = Math.max(this.noiseFloor, 0.00005);
+    // FIXED noise floor with minimal adaptation
+    // Only adapt within strict bounds to prevent drift
+    const targetFloor = Math.max(0.0003, Math.min(0.003, instantRms));
+    this.noiseFloor = this.noiseFloor * 0.995 + targetFloor * 0.005;
+    
+    // Hard clamp to prevent any drift outside acceptable range
+    this.noiseFloor = Math.max(0.0003, Math.min(0.003, this.noiseFloor));
 
-    // SECOND PASS: Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
+    // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
     const snr = this.energy / (this.noiseFloor + 1e-6);
     const snrDb = 20 * Math.log10(Math.max(1e-6, snr));
     
@@ -79,11 +67,20 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     // Probability scales linearly from 0 to 1 between minSNR and (minSNR + snrRange)
     let probability = Math.min(1, Math.max(0, (snrDb - this.minSNR) / this.snrRange));
 
-    // Apply absolute energy threshold
-    // We use a soft threshold to avoid abrupt cutting
+    // Apply absolute energy threshold with soft knee
     if (this.energy < this.minEnergy) {
       const energyRatio = this.energy / (this.minEnergy + 1e-6);
-      probability *= Math.pow(energyRatio, 2); // Quadratic falloff
+      probability *= Math.pow(energyRatio, 2);
+    }
+
+    // Apply crest factor penalty
+    // Reject signals with high crest factor (sharp transients like keyboard clicks)
+    // Voice: 6-12dB, Keyboard: 20-30dB
+    // We penalize anything above 14dB
+    if (crestFactorDb > 14) {
+      const excess = crestFactorDb - 14;
+      const penalty = Math.max(0, 1 - (excess / 10)); // Linear falloff over 10dB
+      probability *= penalty;
     }
 
     this.port.postMessage({ probability, snr: snrDb, noiseFloor: this.noiseFloor, rms: this.energy });

package/dist/{chunk-DLLK6K76.mjs → chunk-KLBA2CPE.mjs}

@@ -20,17 +20,15 @@ var VADStateMachine = class {
       // Smooth for natural speech
       preRollMs: config?.preRollMs ?? 250,
       // Generous pre-roll
-      minSpeechDurationMs: config?.minSpeechDurationMs ?? 150,
-      // Increased to filter keyboard clicks
+      minSpeechDurationMs: config?.minSpeechDurationMs ?? 250,
+      // Aggressive transient rejection
       minSilenceDurationMs: config?.minSilenceDurationMs ?? 150,
       energyVad: {
         smoothing: config?.energyVad?.smoothing ?? 0.95,
         initialNoiseFloor: config?.energyVad?.initialNoiseFloor ?? 1e-3,
-        noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 5e-3,
-        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.1,
-        minSNR: config?.energyVad?.minSNR ?? 10,
-        snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
+        minSNR: config?.energyVad?.minSNR ?? 8,
+        snrRange: config?.energyVad?.snrRange ?? 12,
+        minEnergy: config?.energyVad?.minEnergy ?? 0.01
       }
     };
     this.lastSilenceTime = Date.now();

package/dist/{chunk-K6X52R7N.mjs → chunk-QQFKHTCQ.mjs}

@@ -1,6 +1,6 @@
 import {
   createAudioPipeline
-} from "./chunk-RD4GDIPO.mjs";
+} from "./chunk-WQVMSR7V.mjs";
 
 // src/livekit/integration.ts
 async function attachProcessingToTrack(track, config = {}) {

package/dist/{chunk-OXV7BHX5.mjs → chunk-U26F3GJN.mjs}

@@ -3,7 +3,7 @@ import {
 } from "./chunk-XO6B3D4A.mjs";
 import {
   EnergyVADPlugin
-} from "./chunk-FKR6NWZF.mjs";
+} from "./chunk-GLKAWCEW.mjs";
 
 // src/extensibility/plugins.ts
 var nsPlugins = /* @__PURE__ */ new Map();

package/dist/{chunk-RD4GDIPO.mjs → chunk-WQVMSR7V.mjs}

@@ -1,6 +1,6 @@
 import {
   VADStateMachine
-} from "./chunk-DLLK6K76.mjs";
+} from "./chunk-KLBA2CPE.mjs";
 import {
   getAudioContext,
   registerPipeline,
@@ -9,7 +9,7 @@ import {
 import {
   getNoiseSuppressionPlugin,
   getVADPlugin
-} from "./chunk-OXV7BHX5.mjs";
+} from "./chunk-U26F3GJN.mjs";
 
 // src/pipeline/audio-pipeline.ts
 import mitt from "mitt";
@@ -37,10 +37,9 @@ async function createAudioPipeline(sourceTrack, config = {}) {
       energyVad: {
         smoothing: 0.95,
         initialNoiseFloor: 1e-3,
-        noiseFloorAdaptRateQuiet: 0.01,
-        noiseFloorAdaptRateLoud: 1e-3,
-        minSNR: 2,
-        snrRange: 8
+        minSNR: 8,
+        snrRange: 12,
+        minEnergy: 0.01
       },
       ...config.vad
     },

package/dist/extensibility/plugins.js

@@ -106,11 +106,9 @@ var createEnergyVadWorkletCode = (vadConfig) => {
   const energyParams = vadConfig?.energyVad || {};
   const smoothing = energyParams.smoothing ?? 0.95;
   const initialNoiseFloor = energyParams.initialNoiseFloor ?? 1e-3;
-  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.01;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.1;
-  const minSNR = energyParams.minSNR ?? 10;
-  const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 1e-3;
+  const minSNR = energyParams.minSNR ?? 8;
+  const snrRange = energyParams.snrRange ?? 12;
+  const minEnergy = energyParams.minEnergy ?? 0.01;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -118,8 +116,6 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     this.smoothing = ${smoothing};
     this.energy = 0;
     this.noiseFloor = ${initialNoiseFloor};
-    this.noiseFloorAdaptRateQuiet = ${noiseFloorAdaptRateQuiet};
-    this.noiseFloorAdaptRateLoud = ${noiseFloorAdaptRateLoud};
     this.minSNR = ${minSNR};
     this.snrRange = ${snrRange};
     this.minEnergy = ${minEnergy};
@@ -139,8 +135,11 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     
     // Calculate instantaneous RMS (Root Mean Square) energy
     let sum = 0;
+    let peak = 0;
     for (let i = 0; i < channel.length; i++) {
+      const sample = Math.abs(channel[i]);
       sum += channel[i] * channel[i];
+      peak = Math.max(peak, sample);
     }
     const instantRms = Math.sqrt(sum / channel.length);
 
@@ -148,32 +147,21 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     // this.energy acts as the smoothed RMS value
     this.energy = this.energy * this.smoothing + instantRms * (1 - this.smoothing);
 
-    // Adaptive noise floor estimation
-    // We use a TWO-PASS approach to avoid circular dependencies:
-    // FIRST PASS: Calculate instantaneous SNR to decide how to adapt
-    const instantSnr = instantRms / (this.noiseFloor + 1e-6);
-    const instantSnrDb = 20 * Math.log10(Math.max(1e-6, instantSnr));
-    
-    // Adapt the noise floor based on instantaneous SNR
-    if (instantRms < this.noiseFloor) {
-      // Signal is quieter than noise floor, adapt downwards quickly
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
-    } else if (instantSnrDb < 12) {
-      // Signal is louder but SNR is low (< 12dB) - likely just louder background noise
-      // Adapt upwards at normal rate to track rising noise
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateLoud) + instantRms * this.noiseFloorAdaptRateLoud;
-    } else {
-      // Signal has high SNR (>= 12dB) - likely speech or transient
-      // Adapt VERY slowly to avoid "chasing" speech
-      const slowRate = this.noiseFloorAdaptRateLoud * 0.02;
-      this.noiseFloor = this.noiseFloor * (1 - slowRate) + instantRms * slowRate;
-    }
+    // Calculate Crest Factor (peak-to-RMS ratio)
+    // Voice typically has crest factor of 2-4 (6-12dB)
+    // Keyboard clicks have crest factor of 10-30+ (20-30dB)
+    const crestFactor = peak / (instantRms + 1e-10);
+    const crestFactorDb = 20 * Math.log10(Math.max(1e-6, crestFactor));
 
-    // Ensure noise floor doesn't drop to absolute zero
-    // 0.00005 is approx -86dB, very quiet but prevents SNR explosion
-    this.noiseFloor = Math.max(this.noiseFloor, 0.00005);
+    // FIXED noise floor with minimal adaptation
+    // Only adapt within strict bounds to prevent drift
+    const targetFloor = Math.max(0.0003, Math.min(0.003, instantRms));
+    this.noiseFloor = this.noiseFloor * 0.995 + targetFloor * 0.005;
+    
+    // Hard clamp to prevent any drift outside acceptable range
+    this.noiseFloor = Math.max(0.0003, Math.min(0.003, this.noiseFloor));
 
-    // SECOND PASS: Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
+    // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
     const snr = this.energy / (this.noiseFloor + 1e-6);
     const snrDb = 20 * Math.log10(Math.max(1e-6, snr));
     
@@ -182,11 +170,20 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     // Probability scales linearly from 0 to 1 between minSNR and (minSNR + snrRange)
     let probability = Math.min(1, Math.max(0, (snrDb - this.minSNR) / this.snrRange));
 
-    // Apply absolute energy threshold
-    // We use a soft threshold to avoid abrupt cutting
+    // Apply absolute energy threshold with soft knee
     if (this.energy < this.minEnergy) {
       const energyRatio = this.energy / (this.minEnergy + 1e-6);
-      probability *= Math.pow(energyRatio, 2); // Quadratic falloff
+      probability *= Math.pow(energyRatio, 2);
+    }
+
+    // Apply crest factor penalty
+    // Reject signals with high crest factor (sharp transients like keyboard clicks)
+    // Voice: 6-12dB, Keyboard: 20-30dB
+    // We penalize anything above 14dB
+    if (crestFactorDb > 14) {
+      const excess = crestFactorDb - 14;
+      const penalty = Math.max(0, 1 - (excess / 10)); // Linear falloff over 10dB
+      probability *= penalty;
     }
 
     this.port.postMessage({ probability, snr: snrDb, noiseFloor: this.noiseFloor, rms: this.energy });

package/dist/extensibility/plugins.mjs

@@ -3,9 +3,9 @@ import {
   getVADPlugin,
   registerNoiseSuppressionPlugin,
   registerVADPlugin
-} from "../chunk-OXV7BHX5.mjs";
+} from "../chunk-U26F3GJN.mjs";
 import "../chunk-XO6B3D4A.mjs";
-import "../chunk-FKR6NWZF.mjs";
+import "../chunk-GLKAWCEW.mjs";
 export {
   getNoiseSuppressionPlugin,
   getVADPlugin,

package/dist/index.js

@@ -158,11 +158,9 @@ var createEnergyVadWorkletCode = (vadConfig) => {
   const energyParams = vadConfig?.energyVad || {};
   const smoothing = energyParams.smoothing ?? 0.95;
   const initialNoiseFloor = energyParams.initialNoiseFloor ?? 1e-3;
-  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.01;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.1;
-  const minSNR = energyParams.minSNR ?? 10;
-  const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 1e-3;
+  const minSNR = energyParams.minSNR ?? 8;
+  const snrRange = energyParams.snrRange ?? 12;
+  const minEnergy = energyParams.minEnergy ?? 0.01;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -170,8 +168,6 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     this.smoothing = ${smoothing};
     this.energy = 0;
     this.noiseFloor = ${initialNoiseFloor};
-    this.noiseFloorAdaptRateQuiet = ${noiseFloorAdaptRateQuiet};
-    this.noiseFloorAdaptRateLoud = ${noiseFloorAdaptRateLoud};
     this.minSNR = ${minSNR};
     this.snrRange = ${snrRange};
     this.minEnergy = ${minEnergy};
@@ -191,8 +187,11 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     
     // Calculate instantaneous RMS (Root Mean Square) energy
     let sum = 0;
+    let peak = 0;
     for (let i = 0; i < channel.length; i++) {
+      const sample = Math.abs(channel[i]);
       sum += channel[i] * channel[i];
+      peak = Math.max(peak, sample);
     }
     const instantRms = Math.sqrt(sum / channel.length);
 
@@ -200,32 +199,21 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     // this.energy acts as the smoothed RMS value
     this.energy = this.energy * this.smoothing + instantRms * (1 - this.smoothing);
 
-    // Adaptive noise floor estimation
-    // We use a TWO-PASS approach to avoid circular dependencies:
-    // FIRST PASS: Calculate instantaneous SNR to decide how to adapt
-    const instantSnr = instantRms / (this.noiseFloor + 1e-6);
-    const instantSnrDb = 20 * Math.log10(Math.max(1e-6, instantSnr));
-    
-    // Adapt the noise floor based on instantaneous SNR
-    if (instantRms < this.noiseFloor) {
-      // Signal is quieter than noise floor, adapt downwards quickly
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
-    } else if (instantSnrDb < 12) {
-      // Signal is louder but SNR is low (< 12dB) - likely just louder background noise
-      // Adapt upwards at normal rate to track rising noise
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateLoud) + instantRms * this.noiseFloorAdaptRateLoud;
-    } else {
-      // Signal has high SNR (>= 12dB) - likely speech or transient
-      // Adapt VERY slowly to avoid "chasing" speech
-      const slowRate = this.noiseFloorAdaptRateLoud * 0.02;
-      this.noiseFloor = this.noiseFloor * (1 - slowRate) + instantRms * slowRate;
-    }
+    // Calculate Crest Factor (peak-to-RMS ratio)
+    // Voice typically has crest factor of 2-4 (6-12dB)
+    // Keyboard clicks have crest factor of 10-30+ (20-30dB)
+    const crestFactor = peak / (instantRms + 1e-10);
+    const crestFactorDb = 20 * Math.log10(Math.max(1e-6, crestFactor));
 
-    // Ensure noise floor doesn't drop to absolute zero
-    // 0.00005 is approx -86dB, very quiet but prevents SNR explosion
-    this.noiseFloor = Math.max(this.noiseFloor, 0.00005);
+    // FIXED noise floor with minimal adaptation
+    // Only adapt within strict bounds to prevent drift
+    const targetFloor = Math.max(0.0003, Math.min(0.003, instantRms));
+    this.noiseFloor = this.noiseFloor * 0.995 + targetFloor * 0.005;
+    
+    // Hard clamp to prevent any drift outside acceptable range
+    this.noiseFloor = Math.max(0.0003, Math.min(0.003, this.noiseFloor));
 
-    // SECOND PASS: Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
+    // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
     const snr = this.energy / (this.noiseFloor + 1e-6);
     const snrDb = 20 * Math.log10(Math.max(1e-6, snr));
     
@@ -234,11 +222,20 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     // Probability scales linearly from 0 to 1 between minSNR and (minSNR + snrRange)
     let probability = Math.min(1, Math.max(0, (snrDb - this.minSNR) / this.snrRange));
 
-    // Apply absolute energy threshold
-    // We use a soft threshold to avoid abrupt cutting
+    // Apply absolute energy threshold with soft knee
     if (this.energy < this.minEnergy) {
       const energyRatio = this.energy / (this.minEnergy + 1e-6);
-      probability *= Math.pow(energyRatio, 2); // Quadratic falloff
+      probability *= Math.pow(energyRatio, 2);
+    }
+
+    // Apply crest factor penalty
+    // Reject signals with high crest factor (sharp transients like keyboard clicks)
+    // Voice: 6-12dB, Keyboard: 20-30dB
+    // We penalize anything above 14dB
+    if (crestFactorDb > 14) {
+      const excess = crestFactorDb - 14;
+      const penalty = Math.max(0, 1 - (excess / 10)); // Linear falloff over 10dB
+      probability *= penalty;
     }
 
     this.port.postMessage({ probability, snr: snrDb, noiseFloor: this.noiseFloor, rms: this.energy });
@@ -367,17 +364,15 @@ var VADStateMachine = class {
       // Smooth for natural speech
       preRollMs: config?.preRollMs ?? 250,
       // Generous pre-roll
-      minSpeechDurationMs: config?.minSpeechDurationMs ?? 150,
-      // Increased to filter keyboard clicks
+      minSpeechDurationMs: config?.minSpeechDurationMs ?? 250,
+      // Aggressive transient rejection
       minSilenceDurationMs: config?.minSilenceDurationMs ?? 150,
       energyVad: {
         smoothing: config?.energyVad?.smoothing ?? 0.95,
         initialNoiseFloor: config?.energyVad?.initialNoiseFloor ?? 1e-3,
-        noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 5e-3,
-        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.1,
-        minSNR: config?.energyVad?.minSNR ?? 10,
-        snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
+        minSNR: config?.energyVad?.minSNR ?? 8,
+        snrRange: config?.energyVad?.snrRange ?? 12,
+        minEnergy: config?.energyVad?.minEnergy ?? 0.01
       }
     };
     this.lastSilenceTime = Date.now();
@@ -470,10 +465,9 @@ async function createAudioPipeline(sourceTrack, config = {}) {
       energyVad: {
         smoothing: 0.95,
         initialNoiseFloor: 1e-3,
-        noiseFloorAdaptRateQuiet: 0.01,
-        noiseFloorAdaptRateLoud: 1e-3,
-        minSNR: 2,
-        snrRange: 8
+        minSNR: 8,
+        snrRange: 12,
+        minEnergy: 0.01
       },
       ...config.vad
     },

package/dist/index.mjs

@@ -1,13 +1,13 @@
 import "./chunk-WBQAMGXK.mjs";
 import {
   attachProcessingToTrack
-} from "./chunk-K6X52R7N.mjs";
+} from "./chunk-QQFKHTCQ.mjs";
 import {
   createAudioPipeline
-} from "./chunk-RD4GDIPO.mjs";
+} from "./chunk-WQVMSR7V.mjs";
 import {
   VADStateMachine
-} from "./chunk-DLLK6K76.mjs";
+} from "./chunk-KLBA2CPE.mjs";
 import {
   closeAudioContext,
   getAudioContext,
@@ -21,13 +21,13 @@ import {
   getVADPlugin,
   registerNoiseSuppressionPlugin,
   registerVADPlugin
-} from "./chunk-OXV7BHX5.mjs";
+} from "./chunk-U26F3GJN.mjs";
 import {
   RNNoisePlugin
 } from "./chunk-XO6B3D4A.mjs";
 import {
   EnergyVADPlugin
-} from "./chunk-FKR6NWZF.mjs";
+} from "./chunk-GLKAWCEW.mjs";
 export {
   EnergyVADPlugin,
   RNNoisePlugin,

package/dist/livekit/integration.js

@@ -127,11 +127,9 @@ var createEnergyVadWorkletCode = (vadConfig) => {
   const energyParams = vadConfig?.energyVad || {};
   const smoothing = energyParams.smoothing ?? 0.95;
   const initialNoiseFloor = energyParams.initialNoiseFloor ?? 1e-3;
-  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.01;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.1;
-  const minSNR = energyParams.minSNR ?? 10;
-  const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 1e-3;
+  const minSNR = energyParams.minSNR ?? 8;
+  const snrRange = energyParams.snrRange ?? 12;
+  const minEnergy = energyParams.minEnergy ?? 0.01;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -139,8 +137,6 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     this.smoothing = ${smoothing};
     this.energy = 0;
     this.noiseFloor = ${initialNoiseFloor};
-    this.noiseFloorAdaptRateQuiet = ${noiseFloorAdaptRateQuiet};
-    this.noiseFloorAdaptRateLoud = ${noiseFloorAdaptRateLoud};
     this.minSNR = ${minSNR};
     this.snrRange = ${snrRange};
     this.minEnergy = ${minEnergy};
@@ -160,8 +156,11 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     
     // Calculate instantaneous RMS (Root Mean Square) energy
     let sum = 0;
+    let peak = 0;
     for (let i = 0; i < channel.length; i++) {
+      const sample = Math.abs(channel[i]);
       sum += channel[i] * channel[i];
+      peak = Math.max(peak, sample);
     }
     const instantRms = Math.sqrt(sum / channel.length);
 
@@ -169,32 +168,21 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     // this.energy acts as the smoothed RMS value
     this.energy = this.energy * this.smoothing + instantRms * (1 - this.smoothing);
 
-    // Adaptive noise floor estimation
-    // We use a TWO-PASS approach to avoid circular dependencies:
-    // FIRST PASS: Calculate instantaneous SNR to decide how to adapt
-    const instantSnr = instantRms / (this.noiseFloor + 1e-6);
-    const instantSnrDb = 20 * Math.log10(Math.max(1e-6, instantSnr));
-    
-    // Adapt the noise floor based on instantaneous SNR
-    if (instantRms < this.noiseFloor) {
-      // Signal is quieter than noise floor, adapt downwards quickly
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
-    } else if (instantSnrDb < 12) {
-      // Signal is louder but SNR is low (< 12dB) - likely just louder background noise
-      // Adapt upwards at normal rate to track rising noise
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateLoud) + instantRms * this.noiseFloorAdaptRateLoud;
-    } else {
-      // Signal has high SNR (>= 12dB) - likely speech or transient
-      // Adapt VERY slowly to avoid "chasing" speech
-      const slowRate = this.noiseFloorAdaptRateLoud * 0.02;
-      this.noiseFloor = this.noiseFloor * (1 - slowRate) + instantRms * slowRate;
-    }
+    // Calculate Crest Factor (peak-to-RMS ratio)
+    // Voice typically has crest factor of 2-4 (6-12dB)
+    // Keyboard clicks have crest factor of 10-30+ (20-30dB)
+    const crestFactor = peak / (instantRms + 1e-10);
+    const crestFactorDb = 20 * Math.log10(Math.max(1e-6, crestFactor));
 
-    // Ensure noise floor doesn't drop to absolute zero
-    // 0.00005 is approx -86dB, very quiet but prevents SNR explosion
-    this.noiseFloor = Math.max(this.noiseFloor, 0.00005);
+    // FIXED noise floor with minimal adaptation
+    // Only adapt within strict bounds to prevent drift
+    const targetFloor = Math.max(0.0003, Math.min(0.003, instantRms));
+    this.noiseFloor = this.noiseFloor * 0.995 + targetFloor * 0.005;
+    
+    // Hard clamp to prevent any drift outside acceptable range
+    this.noiseFloor = Math.max(0.0003, Math.min(0.003, this.noiseFloor));
 
-    // SECOND PASS: Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
+    // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
     const snr = this.energy / (this.noiseFloor + 1e-6);
     const snrDb = 20 * Math.log10(Math.max(1e-6, snr));
     
@@ -203,11 +191,20 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     // Probability scales linearly from 0 to 1 between minSNR and (minSNR + snrRange)
     let probability = Math.min(1, Math.max(0, (snrDb - this.minSNR) / this.snrRange));
 
-    // Apply absolute energy threshold
-    // We use a soft threshold to avoid abrupt cutting
+    // Apply absolute energy threshold with soft knee
     if (this.energy < this.minEnergy) {
       const energyRatio = this.energy / (this.minEnergy + 1e-6);
-      probability *= Math.pow(energyRatio, 2); // Quadratic falloff
+      probability *= Math.pow(energyRatio, 2);
+    }
+
+    // Apply crest factor penalty
+    // Reject signals with high crest factor (sharp transients like keyboard clicks)
+    // Voice: 6-12dB, Keyboard: 20-30dB
+    // We penalize anything above 14dB
+    if (crestFactorDb > 14) {
+      const excess = crestFactorDb - 14;
+      const penalty = Math.max(0, 1 - (excess / 10)); // Linear falloff over 10dB
+      probability *= penalty;
     }
 
     this.port.postMessage({ probability, snr: snrDb, noiseFloor: this.noiseFloor, rms: this.energy });
@@ -330,17 +327,15 @@ var VADStateMachine = class {
       // Smooth for natural speech
       preRollMs: config?.preRollMs ?? 250,
       // Generous pre-roll
-      minSpeechDurationMs: config?.minSpeechDurationMs ?? 150,
-      // Increased to filter keyboard clicks
+      minSpeechDurationMs: config?.minSpeechDurationMs ?? 250,
+      // Aggressive transient rejection
       minSilenceDurationMs: config?.minSilenceDurationMs ?? 150,
       energyVad: {
         smoothing: config?.energyVad?.smoothing ?? 0.95,
         initialNoiseFloor: config?.energyVad?.initialNoiseFloor ?? 1e-3,
-        noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 5e-3,
-        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.1,
-        minSNR: config?.energyVad?.minSNR ?? 10,
-        snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
+        minSNR: config?.energyVad?.minSNR ?? 8,
+        snrRange: config?.energyVad?.snrRange ?? 12,
+        minEnergy: config?.energyVad?.minEnergy ?? 0.01
       }
     };
     this.lastSilenceTime = Date.now();
@@ -433,10 +428,9 @@ async function createAudioPipeline(sourceTrack, config = {}) {
       energyVad: {
         smoothing: 0.95,
         initialNoiseFloor: 1e-3,
-        noiseFloorAdaptRateQuiet: 0.01,
-        noiseFloorAdaptRateLoud: 1e-3,
-        minSNR: 2,
-        snrRange: 8
+        minSNR: 8,
+        snrRange: 12,
+        minEnergy: 0.01
       },
       ...config.vad
     },

package/dist/livekit/integration.mjs

@@ -1,12 +1,12 @@
 import {
   attachProcessingToTrack
-} from "../chunk-K6X52R7N.mjs";
-import "../chunk-RD4GDIPO.mjs";
-import "../chunk-DLLK6K76.mjs";
+} from "../chunk-QQFKHTCQ.mjs";
+import "../chunk-WQVMSR7V.mjs";
+import "../chunk-KLBA2CPE.mjs";
 import "../chunk-OZ7KMC4S.mjs";
-import "../chunk-OXV7BHX5.mjs";
+import "../chunk-U26F3GJN.mjs";
 import "../chunk-XO6B3D4A.mjs";
-import "../chunk-FKR6NWZF.mjs";
+import "../chunk-GLKAWCEW.mjs";
 export {
   attachProcessingToTrack
 };

package/dist/pipeline/audio-pipeline.js

@@ -125,11 +125,9 @@ var createEnergyVadWorkletCode = (vadConfig) => {
   const energyParams = vadConfig?.energyVad || {};
   const smoothing = energyParams.smoothing ?? 0.95;
   const initialNoiseFloor = energyParams.initialNoiseFloor ?? 1e-3;
-  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.01;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.1;
-  const minSNR = energyParams.minSNR ?? 10;
-  const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 1e-3;
+  const minSNR = energyParams.minSNR ?? 8;
+  const snrRange = energyParams.snrRange ?? 12;
+  const minEnergy = energyParams.minEnergy ?? 0.01;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -137,8 +135,6 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     this.smoothing = ${smoothing};
     this.energy = 0;
     this.noiseFloor = ${initialNoiseFloor};
-    this.noiseFloorAdaptRateQuiet = ${noiseFloorAdaptRateQuiet};
-    this.noiseFloorAdaptRateLoud = ${noiseFloorAdaptRateLoud};
     this.minSNR = ${minSNR};
     this.snrRange = ${snrRange};
     this.minEnergy = ${minEnergy};
@@ -158,8 +154,11 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     
     // Calculate instantaneous RMS (Root Mean Square) energy
     let sum = 0;
+    let peak = 0;
     for (let i = 0; i < channel.length; i++) {
+      const sample = Math.abs(channel[i]);
       sum += channel[i] * channel[i];
+      peak = Math.max(peak, sample);
     }
     const instantRms = Math.sqrt(sum / channel.length);
 
@@ -167,32 +166,21 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     // this.energy acts as the smoothed RMS value
     this.energy = this.energy * this.smoothing + instantRms * (1 - this.smoothing);
 
-    // Adaptive noise floor estimation
-    // We use a TWO-PASS approach to avoid circular dependencies:
-    // FIRST PASS: Calculate instantaneous SNR to decide how to adapt
-    const instantSnr = instantRms / (this.noiseFloor + 1e-6);
-    const instantSnrDb = 20 * Math.log10(Math.max(1e-6, instantSnr));
-    
-    // Adapt the noise floor based on instantaneous SNR
-    if (instantRms < this.noiseFloor) {
-      // Signal is quieter than noise floor, adapt downwards quickly
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
-    } else if (instantSnrDb < 12) {
-      // Signal is louder but SNR is low (< 12dB) - likely just louder background noise
-      // Adapt upwards at normal rate to track rising noise
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateLoud) + instantRms * this.noiseFloorAdaptRateLoud;
-    } else {
-      // Signal has high SNR (>= 12dB) - likely speech or transient
-      // Adapt VERY slowly to avoid "chasing" speech
-      const slowRate = this.noiseFloorAdaptRateLoud * 0.02;
-      this.noiseFloor = this.noiseFloor * (1 - slowRate) + instantRms * slowRate;
-    }
+    // Calculate Crest Factor (peak-to-RMS ratio)
+    // Voice typically has crest factor of 2-4 (6-12dB)
+    // Keyboard clicks have crest factor of 10-30+ (20-30dB)
+    const crestFactor = peak / (instantRms + 1e-10);
+    const crestFactorDb = 20 * Math.log10(Math.max(1e-6, crestFactor));
 
-    // Ensure noise floor doesn't drop to absolute zero
-    // 0.00005 is approx -86dB, very quiet but prevents SNR explosion
-    this.noiseFloor = Math.max(this.noiseFloor, 0.00005);
+    // FIXED noise floor with minimal adaptation
+    // Only adapt within strict bounds to prevent drift
+    const targetFloor = Math.max(0.0003, Math.min(0.003, instantRms));
+    this.noiseFloor = this.noiseFloor * 0.995 + targetFloor * 0.005;
+    
+    // Hard clamp to prevent any drift outside acceptable range
+    this.noiseFloor = Math.max(0.0003, Math.min(0.003, this.noiseFloor));
 
-    // SECOND PASS: Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
+    // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
     const snr = this.energy / (this.noiseFloor + 1e-6);
     const snrDb = 20 * Math.log10(Math.max(1e-6, snr));
     
@@ -201,11 +189,20 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     // Probability scales linearly from 0 to 1 between minSNR and (minSNR + snrRange)
     let probability = Math.min(1, Math.max(0, (snrDb - this.minSNR) / this.snrRange));
 
-    // Apply absolute energy threshold
-    // We use a soft threshold to avoid abrupt cutting
+    // Apply absolute energy threshold with soft knee
     if (this.energy < this.minEnergy) {
       const energyRatio = this.energy / (this.minEnergy + 1e-6);
-      probability *= Math.pow(energyRatio, 2); // Quadratic falloff
+      probability *= Math.pow(energyRatio, 2);
+    }
+
+    // Apply crest factor penalty
+    // Reject signals with high crest factor (sharp transients like keyboard clicks)
+    // Voice: 6-12dB, Keyboard: 20-30dB
+    // We penalize anything above 14dB
+    if (crestFactorDb > 14) {
+      const excess = crestFactorDb - 14;
+      const penalty = Math.max(0, 1 - (excess / 10)); // Linear falloff over 10dB
+      probability *= penalty;
     }
 
     this.port.postMessage({ probability, snr: snrDb, noiseFloor: this.noiseFloor, rms: this.energy });
@@ -328,17 +325,15 @@ var VADStateMachine = class {
       // Smooth for natural speech
       preRollMs: config?.preRollMs ?? 250,
       // Generous pre-roll
-      minSpeechDurationMs: config?.minSpeechDurationMs ?? 150,
-      // Increased to filter keyboard clicks
+      minSpeechDurationMs: config?.minSpeechDurationMs ?? 250,
+      // Aggressive transient rejection
       minSilenceDurationMs: config?.minSilenceDurationMs ?? 150,
       energyVad: {
         smoothing: config?.energyVad?.smoothing ?? 0.95,
         initialNoiseFloor: config?.energyVad?.initialNoiseFloor ?? 1e-3,
-        noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 5e-3,
-        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.1,
-        minSNR: config?.energyVad?.minSNR ?? 10,
-        snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
+        minSNR: config?.energyVad?.minSNR ?? 8,
+        snrRange: config?.energyVad?.snrRange ?? 12,
+        minEnergy: config?.energyVad?.minEnergy ?? 0.01
       }
     };
     this.lastSilenceTime = Date.now();
@@ -431,10 +426,9 @@ async function createAudioPipeline(sourceTrack, config = {}) {
       energyVad: {
         smoothing: 0.95,
         initialNoiseFloor: 1e-3,
-        noiseFloorAdaptRateQuiet: 0.01,
-        noiseFloorAdaptRateLoud: 1e-3,
-        minSNR: 2,
-        snrRange: 8
+        minSNR: 8,
+        snrRange: 12,
+        minEnergy: 0.01
       },
       ...config.vad
     },

package/dist/pipeline/audio-pipeline.mjs

@@ -1,11 +1,11 @@
 import {
   createAudioPipeline
-} from "../chunk-RD4GDIPO.mjs";
-import "../chunk-DLLK6K76.mjs";
+} from "../chunk-WQVMSR7V.mjs";
+import "../chunk-KLBA2CPE.mjs";
 import "../chunk-OZ7KMC4S.mjs";
-import "../chunk-OXV7BHX5.mjs";
+import "../chunk-U26F3GJN.mjs";
 import "../chunk-XO6B3D4A.mjs";
-import "../chunk-FKR6NWZF.mjs";
+import "../chunk-GLKAWCEW.mjs";
 export {
   createAudioPipeline
 };

package/dist/types.d.mts

@@ -70,8 +70,8 @@ interface AudioProcessingConfig {
         preRollMs?: number;
         /**
          * Minimum speech duration in ms to consider it valid speech.
-         * Filters out very brief noise spikes like keyboard clicks.
-         * Default: 150ms
+         * Filters out brief transients like keyboard clicks.
+         * Default: 250ms (aggressive transient rejection)
          */
         minSpeechDurationMs?: number;
         /**
@@ -95,31 +95,20 @@ interface AudioProcessingConfig {
              * Default: 0.001
              */
             initialNoiseFloor?: number;
-            /**
-             * Rate at which noise floor adapts to quiet signals (0-1).
-             * Default: 0.01
-             */
-            noiseFloorAdaptRateQuiet?: number;
-            /**
-             * Rate at which noise floor adapts to loud signals (0-1).
-             * Applied when instantaneous SNR < 12dB (background noise).
-             * Default: 0.1 (fast tracking of rising noise)
-             */
-            noiseFloorAdaptRateLoud?: number;
             /**
              * Minimum SNR (Signal-to-Noise Ratio) in dB for speech detection.
-             * Default: 10.0 (more aggressive noise rejection)
+             * Default: 8.0
              */
             minSNR?: number;
             /**
              * SNR range in dB for probability scaling.
-             * Default: 10.0 (probability scales from minSNR to minSNR+snrRange)
+             * Default: 12.0 (probability scales from minSNR to minSNR+snrRange)
              */
             snrRange?: number;
             /**
              * Minimum absolute RMS energy to consider as speech.
-             * Prevents triggering on very quiet background noise in silent rooms.
-             * Default: 0.001 (approx -60dB)
+             * Prevents triggering on quiet background noise.
+             * Default: 0.01 (approx -40dB, typical voice level)
              */
             minEnergy?: number;
         };

package/dist/types.d.ts

@@ -70,8 +70,8 @@ interface AudioProcessingConfig {
         preRollMs?: number;
         /**
          * Minimum speech duration in ms to consider it valid speech.
-         * Filters out very brief noise spikes like keyboard clicks.
-         * Default: 150ms
+         * Filters out brief transients like keyboard clicks.
+         * Default: 250ms (aggressive transient rejection)
          */
         minSpeechDurationMs?: number;
         /**
@@ -95,31 +95,20 @@ interface AudioProcessingConfig {
              * Default: 0.001
              */
             initialNoiseFloor?: number;
-            /**
-             * Rate at which noise floor adapts to quiet signals (0-1).
-             * Default: 0.01
-             */
-            noiseFloorAdaptRateQuiet?: number;
-            /**
-             * Rate at which noise floor adapts to loud signals (0-1).
-             * Applied when instantaneous SNR < 12dB (background noise).
-             * Default: 0.1 (fast tracking of rising noise)
-             */
-            noiseFloorAdaptRateLoud?: number;
             /**
              * Minimum SNR (Signal-to-Noise Ratio) in dB for speech detection.
-             * Default: 10.0 (more aggressive noise rejection)
+             * Default: 8.0
              */
             minSNR?: number;
             /**
              * SNR range in dB for probability scaling.
-             * Default: 10.0 (probability scales from minSNR to minSNR+snrRange)
+             * Default: 12.0 (probability scales from minSNR to minSNR+snrRange)
              */
             snrRange?: number;
             /**
              * Minimum absolute RMS energy to consider as speech.
-             * Prevents triggering on very quiet background noise in silent rooms.
-             * Default: 0.001 (approx -60dB)
+             * Prevents triggering on quiet background noise.
+             * Default: 0.01 (approx -40dB, typical voice level)
              */
             minEnergy?: number;
         };

package/dist/vad/vad-node.js

@@ -27,11 +27,9 @@ var createEnergyVadWorkletCode = (vadConfig) => {
   const energyParams = vadConfig?.energyVad || {};
   const smoothing = energyParams.smoothing ?? 0.95;
   const initialNoiseFloor = energyParams.initialNoiseFloor ?? 1e-3;
-  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.01;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.1;
-  const minSNR = energyParams.minSNR ?? 10;
-  const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 1e-3;
+  const minSNR = energyParams.minSNR ?? 8;
+  const snrRange = energyParams.snrRange ?? 12;
+  const minEnergy = energyParams.minEnergy ?? 0.01;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -39,8 +37,6 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     this.smoothing = ${smoothing};
     this.energy = 0;
     this.noiseFloor = ${initialNoiseFloor};
-    this.noiseFloorAdaptRateQuiet = ${noiseFloorAdaptRateQuiet};
-    this.noiseFloorAdaptRateLoud = ${noiseFloorAdaptRateLoud};
     this.minSNR = ${minSNR};
     this.snrRange = ${snrRange};
     this.minEnergy = ${minEnergy};
@@ -60,8 +56,11 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     
     // Calculate instantaneous RMS (Root Mean Square) energy
     let sum = 0;
+    let peak = 0;
     for (let i = 0; i < channel.length; i++) {
+      const sample = Math.abs(channel[i]);
       sum += channel[i] * channel[i];
+      peak = Math.max(peak, sample);
     }
     const instantRms = Math.sqrt(sum / channel.length);
 
@@ -69,32 +68,21 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     // this.energy acts as the smoothed RMS value
     this.energy = this.energy * this.smoothing + instantRms * (1 - this.smoothing);
 
-    // Adaptive noise floor estimation
-    // We use a TWO-PASS approach to avoid circular dependencies:
-    // FIRST PASS: Calculate instantaneous SNR to decide how to adapt
-    const instantSnr = instantRms / (this.noiseFloor + 1e-6);
-    const instantSnrDb = 20 * Math.log10(Math.max(1e-6, instantSnr));
-    
-    // Adapt the noise floor based on instantaneous SNR
-    if (instantRms < this.noiseFloor) {
-      // Signal is quieter than noise floor, adapt downwards quickly
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
-    } else if (instantSnrDb < 12) {
-      // Signal is louder but SNR is low (< 12dB) - likely just louder background noise
-      // Adapt upwards at normal rate to track rising noise
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateLoud) + instantRms * this.noiseFloorAdaptRateLoud;
-    } else {
-      // Signal has high SNR (>= 12dB) - likely speech or transient
-      // Adapt VERY slowly to avoid "chasing" speech
-      const slowRate = this.noiseFloorAdaptRateLoud * 0.02;
-      this.noiseFloor = this.noiseFloor * (1 - slowRate) + instantRms * slowRate;
-    }
+    // Calculate Crest Factor (peak-to-RMS ratio)
+    // Voice typically has crest factor of 2-4 (6-12dB)
+    // Keyboard clicks have crest factor of 10-30+ (20-30dB)
+    const crestFactor = peak / (instantRms + 1e-10);
+    const crestFactorDb = 20 * Math.log10(Math.max(1e-6, crestFactor));
 
-    // Ensure noise floor doesn't drop to absolute zero
-    // 0.00005 is approx -86dB, very quiet but prevents SNR explosion
-    this.noiseFloor = Math.max(this.noiseFloor, 0.00005);
+    // FIXED noise floor with minimal adaptation
+    // Only adapt within strict bounds to prevent drift
+    const targetFloor = Math.max(0.0003, Math.min(0.003, instantRms));
+    this.noiseFloor = this.noiseFloor * 0.995 + targetFloor * 0.005;
+    
+    // Hard clamp to prevent any drift outside acceptable range
+    this.noiseFloor = Math.max(0.0003, Math.min(0.003, this.noiseFloor));
 
-    // SECOND PASS: Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
+    // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
     const snr = this.energy / (this.noiseFloor + 1e-6);
     const snrDb = 20 * Math.log10(Math.max(1e-6, snr));
     
@@ -103,11 +91,20 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     // Probability scales linearly from 0 to 1 between minSNR and (minSNR + snrRange)
     let probability = Math.min(1, Math.max(0, (snrDb - this.minSNR) / this.snrRange));
 
-    // Apply absolute energy threshold
-    // We use a soft threshold to avoid abrupt cutting
+    // Apply absolute energy threshold with soft knee
     if (this.energy < this.minEnergy) {
       const energyRatio = this.energy / (this.minEnergy + 1e-6);
-      probability *= Math.pow(energyRatio, 2); // Quadratic falloff
+      probability *= Math.pow(energyRatio, 2);
+    }
+
+    // Apply crest factor penalty
+    // Reject signals with high crest factor (sharp transients like keyboard clicks)
+    // Voice: 6-12dB, Keyboard: 20-30dB
+    // We penalize anything above 14dB
+    if (crestFactorDb > 14) {
+      const excess = crestFactorDb - 14;
+      const penalty = Math.max(0, 1 - (excess / 10)); // Linear falloff over 10dB
+      probability *= penalty;
     }
 
     this.port.postMessage({ probability, snr: snrDb, noiseFloor: this.noiseFloor, rms: this.energy });

package/dist/vad/vad-node.mjs

@@ -1,6 +1,6 @@
 import {
   EnergyVADPlugin
-} from "../chunk-FKR6NWZF.mjs";
+} from "../chunk-GLKAWCEW.mjs";
 export {
   EnergyVADPlugin
 };

package/dist/vad/vad-state.js

@@ -44,17 +44,15 @@ var VADStateMachine = class {
       // Smooth for natural speech
       preRollMs: config?.preRollMs ?? 250,
       // Generous pre-roll
-      minSpeechDurationMs: config?.minSpeechDurationMs ?? 150,
-      // Increased to filter keyboard clicks
+      minSpeechDurationMs: config?.minSpeechDurationMs ?? 250,
+      // Aggressive transient rejection
       minSilenceDurationMs: config?.minSilenceDurationMs ?? 150,
       energyVad: {
         smoothing: config?.energyVad?.smoothing ?? 0.95,
         initialNoiseFloor: config?.energyVad?.initialNoiseFloor ?? 1e-3,
-        noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 5e-3,
-        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.1,
-        minSNR: config?.energyVad?.minSNR ?? 10,
-        snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
+        minSNR: config?.energyVad?.minSNR ?? 8,
+        snrRange: config?.energyVad?.snrRange ?? 12,
+        minEnergy: config?.energyVad?.minEnergy ?? 0.01
       }
     };
     this.lastSilenceTime = Date.now();

package/dist/vad/vad-state.mjs

@@ -1,6 +1,6 @@
 import {
   VADStateMachine
-} from "../chunk-DLLK6K76.mjs";
+} from "../chunk-KLBA2CPE.mjs";
 export {
   VADStateMachine
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tensamin/audio",
-  "version": "0.1.13",
+  "version": "0.1.15",
   "main": "dist/index.js",
   "module": "dist/index.mjs",
   "types": "dist/index.d.ts",