@tensamin/audio 0.1.12 → 0.1.14

package/README.md

@@ -101,11 +101,11 @@ vad: {
   energyVad?: {
     smoothing: number;                 // Default: 0.95
     initialNoiseFloor: number;         // Default: 0.001
-    noiseFloorAdaptRateQuiet: number;  // Default: 0.01
-    noiseFloorAdaptRateLoud: number;   // Default: 0.05
-    minSNR: number;                    // Default: 10.0 (dB)
+    noiseFloorAdaptRateQuiet: number;  // Default: 0.002
+    noiseFloorAdaptRateLoud: number;   // Default: 0.02
+    minSNR: number;                    // Default: 12.0 (dB)
     snrRange: number;                  // Default: 10.0 (dB)
-    minEnergy: number;                 // Default: 0.001
+    minEnergy: number;                 // Default: 0.003
   };
 }
 ```
@@ -116,7 +116,7 @@ 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 +124,7 @@ 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: 0.003, ~-50dB)
 
 ### Output Control
 

package/dist/{chunk-KEWK2OKV.mjs → chunk-2G2JFHJY.mjs}

@@ -3,11 +3,11 @@ 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 ?? 0.05;
-  const minSNR = energyParams.minSNR ?? 10;
+  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 2e-3;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.02;
+  const minSNR = energyParams.minSNR ?? 12;
   const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 1e-3;
+  const minEnergy = energyParams.minEnergy ?? 3e-3;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -36,8 +36,11 @@ 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);
 
@@ -45,36 +48,39 @@ 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 the instantaneous RMS for noise floor tracking to react quickly to silence
-    if (instantRms < this.noiseFloor) {
-      // If signal is quieter than noise floor, adapt downwards quickly
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
+    // 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));
+
+    // Adaptive noise floor estimation using SMOOTHED energy (not instantaneous)
+    // This prevents sharp transients from affecting the noise floor
+    if (this.energy < this.noiseFloor) {
+      // Signal is quieter than noise floor, adapt downwards slowly
+      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + this.energy * this.noiseFloorAdaptRateQuiet;
     } else {
-      // If signal is louder, adapt upwards
-      // We use a multi-stage adaptation rate:
-      // 1. If we are officially speaking, adapt EXTREMELY slowly (0.01x)
-      // 2. If SNR is very high (> 20dB), assume it's speech and adapt very slowly (0.1x)
-      // 3. Otherwise, adapt at the normal loud rate
-      const snr = instantRms / (this.noiseFloor + 1e-6);
-      const snrDb = 20 * Math.log10(Math.max(1e-6, snr));
+      // Calculate SNR based on smoothed energy
+      const smoothedSnr = this.energy / (this.noiseFloor + 1e-6);
+      const smoothedSnrDb = 20 * Math.log10(Math.max(1e-6, smoothedSnr));
       
-      let multiplier = 1.0;
-      if (this.isSpeaking) {
-        multiplier = 0.05;
-      } else if (snrDb > 20) {
-        multiplier = 0.2;
+      // Only adapt upwards if:
+      // 1. SNR is low (< 10dB) - likely just background noise
+      // 2. AND crest factor is low (< 15dB) - not a sharp transient
+      if (smoothedSnrDb < 10 && crestFactorDb < 15) {
+        // This is persistent background noise, adapt upwards
+        this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateLoud) + this.energy * this.noiseFloorAdaptRateLoud;
+      } else {
+        // Either high SNR (speech) or high crest factor (click) - adapt very slowly
+        const slowRate = this.noiseFloorAdaptRateLoud * 0.01;
+        this.noiseFloor = this.noiseFloor * (1 - slowRate) + this.energy * slowRate;
       }
-      
-      const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
-      this.noiseFloor = this.noiseFloor * (1 - adaptRate) + instantRms * adaptRate;
     }
 
     // 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);
+    this.noiseFloor = Math.max(this.noiseFloor, 0.0001);
 
-    // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
+    // SECOND PASS: 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));
     
@@ -83,11 +89,20 @@ 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-Q2I22TJG.mjs → chunk-6F2HZUYO.mjs}

@@ -1,6 +1,6 @@
 import {
   createAudioPipeline
-} from "./chunk-SMZJFNRU.mjs";
+} from "./chunk-R5M2DGAQ.mjs";
 
 // src/livekit/integration.ts
 async function attachProcessingToTrack(track, config = {}) {

package/dist/{chunk-DYY2MXMU.mjs → chunk-K4YLH73B.mjs}

@@ -20,17 +20,17 @@ 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 ?? 0.01,
-        minSNR: config?.energyVad?.minSNR ?? 10,
+        noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 2e-3,
+        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.02,
+        minSNR: config?.energyVad?.minSNR ?? 12,
         snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
+        minEnergy: config?.energyVad?.minEnergy ?? 3e-3
       }
     };
     this.lastSilenceTime = Date.now();

package/dist/{chunk-SMZJFNRU.mjs → chunk-R5M2DGAQ.mjs}

@@ -1,6 +1,6 @@
 import {
   VADStateMachine
-} from "./chunk-DYY2MXMU.mjs";
+} from "./chunk-K4YLH73B.mjs";
 import {
   getAudioContext,
   registerPipeline,
@@ -9,7 +9,7 @@ import {
 import {
   getNoiseSuppressionPlugin,
   getVADPlugin
-} from "./chunk-XZSFQJW4.mjs";
+} from "./chunk-UFKIAMG3.mjs";
 
 // src/pipeline/audio-pipeline.ts
 import mitt from "mitt";

package/dist/{chunk-XZSFQJW4.mjs → chunk-UFKIAMG3.mjs}

@@ -3,7 +3,7 @@ import {
 } from "./chunk-XO6B3D4A.mjs";
 import {
   EnergyVADPlugin
-} from "./chunk-KEWK2OKV.mjs";
+} from "./chunk-2G2JFHJY.mjs";
 
 // src/extensibility/plugins.ts
 var nsPlugins = /* @__PURE__ */ new Map();

package/dist/extensibility/plugins.js

@@ -106,11 +106,11 @@ 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 ?? 0.05;
-  const minSNR = energyParams.minSNR ?? 10;
+  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 2e-3;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.02;
+  const minSNR = energyParams.minSNR ?? 12;
   const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 1e-3;
+  const minEnergy = energyParams.minEnergy ?? 3e-3;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -139,8 +139,11 @@ 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);
 
@@ -148,36 +151,39 @@ 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 the instantaneous RMS for noise floor tracking to react quickly to silence
-    if (instantRms < this.noiseFloor) {
-      // If signal is quieter than noise floor, adapt downwards quickly
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
+    // 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));
+
+    // Adaptive noise floor estimation using SMOOTHED energy (not instantaneous)
+    // This prevents sharp transients from affecting the noise floor
+    if (this.energy < this.noiseFloor) {
+      // Signal is quieter than noise floor, adapt downwards slowly
+      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + this.energy * this.noiseFloorAdaptRateQuiet;
     } else {
-      // If signal is louder, adapt upwards
-      // We use a multi-stage adaptation rate:
-      // 1. If we are officially speaking, adapt EXTREMELY slowly (0.01x)
-      // 2. If SNR is very high (> 20dB), assume it's speech and adapt very slowly (0.1x)
-      // 3. Otherwise, adapt at the normal loud rate
-      const snr = instantRms / (this.noiseFloor + 1e-6);
-      const snrDb = 20 * Math.log10(Math.max(1e-6, snr));
+      // Calculate SNR based on smoothed energy
+      const smoothedSnr = this.energy / (this.noiseFloor + 1e-6);
+      const smoothedSnrDb = 20 * Math.log10(Math.max(1e-6, smoothedSnr));
       
-      let multiplier = 1.0;
-      if (this.isSpeaking) {
-        multiplier = 0.05;
-      } else if (snrDb > 20) {
-        multiplier = 0.2;
+      // Only adapt upwards if:
+      // 1. SNR is low (< 10dB) - likely just background noise
+      // 2. AND crest factor is low (< 15dB) - not a sharp transient
+      if (smoothedSnrDb < 10 && crestFactorDb < 15) {
+        // This is persistent background noise, adapt upwards
+        this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateLoud) + this.energy * this.noiseFloorAdaptRateLoud;
+      } else {
+        // Either high SNR (speech) or high crest factor (click) - adapt very slowly
+        const slowRate = this.noiseFloorAdaptRateLoud * 0.01;
+        this.noiseFloor = this.noiseFloor * (1 - slowRate) + this.energy * slowRate;
       }
-      
-      const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
-      this.noiseFloor = this.noiseFloor * (1 - adaptRate) + instantRms * adaptRate;
     }
 
     // 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);
+    this.noiseFloor = Math.max(this.noiseFloor, 0.0001);
 
-    // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
+    // SECOND PASS: 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));
     
@@ -186,11 +192,20 @@ 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-XZSFQJW4.mjs";
+} from "../chunk-UFKIAMG3.mjs";
 import "../chunk-XO6B3D4A.mjs";
-import "../chunk-KEWK2OKV.mjs";
+import "../chunk-2G2JFHJY.mjs";
 export {
   getNoiseSuppressionPlugin,
   getVADPlugin,

package/dist/index.js

@@ -158,11 +158,11 @@ 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 ?? 0.05;
-  const minSNR = energyParams.minSNR ?? 10;
+  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 2e-3;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.02;
+  const minSNR = energyParams.minSNR ?? 12;
   const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 1e-3;
+  const minEnergy = energyParams.minEnergy ?? 3e-3;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -191,8 +191,11 @@ 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);
 
@@ -200,36 +203,39 @@ 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 the instantaneous RMS for noise floor tracking to react quickly to silence
-    if (instantRms < this.noiseFloor) {
-      // If signal is quieter than noise floor, adapt downwards quickly
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
+    // 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));
+
+    // Adaptive noise floor estimation using SMOOTHED energy (not instantaneous)
+    // This prevents sharp transients from affecting the noise floor
+    if (this.energy < this.noiseFloor) {
+      // Signal is quieter than noise floor, adapt downwards slowly
+      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + this.energy * this.noiseFloorAdaptRateQuiet;
     } else {
-      // If signal is louder, adapt upwards
-      // We use a multi-stage adaptation rate:
-      // 1. If we are officially speaking, adapt EXTREMELY slowly (0.01x)
-      // 2. If SNR is very high (> 20dB), assume it's speech and adapt very slowly (0.1x)
-      // 3. Otherwise, adapt at the normal loud rate
-      const snr = instantRms / (this.noiseFloor + 1e-6);
-      const snrDb = 20 * Math.log10(Math.max(1e-6, snr));
+      // Calculate SNR based on smoothed energy
+      const smoothedSnr = this.energy / (this.noiseFloor + 1e-6);
+      const smoothedSnrDb = 20 * Math.log10(Math.max(1e-6, smoothedSnr));
       
-      let multiplier = 1.0;
-      if (this.isSpeaking) {
-        multiplier = 0.05;
-      } else if (snrDb > 20) {
-        multiplier = 0.2;
+      // Only adapt upwards if:
+      // 1. SNR is low (< 10dB) - likely just background noise
+      // 2. AND crest factor is low (< 15dB) - not a sharp transient
+      if (smoothedSnrDb < 10 && crestFactorDb < 15) {
+        // This is persistent background noise, adapt upwards
+        this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateLoud) + this.energy * this.noiseFloorAdaptRateLoud;
+      } else {
+        // Either high SNR (speech) or high crest factor (click) - adapt very slowly
+        const slowRate = this.noiseFloorAdaptRateLoud * 0.01;
+        this.noiseFloor = this.noiseFloor * (1 - slowRate) + this.energy * slowRate;
       }
-      
-      const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
-      this.noiseFloor = this.noiseFloor * (1 - adaptRate) + instantRms * adaptRate;
     }
 
     // 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);
+    this.noiseFloor = Math.max(this.noiseFloor, 0.0001);
 
-    // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
+    // SECOND PASS: 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));
     
@@ -238,11 +244,20 @@ 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 });
@@ -371,17 +386,17 @@ 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 ?? 0.01,
-        minSNR: config?.energyVad?.minSNR ?? 10,
+        noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 2e-3,
+        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.02,
+        minSNR: config?.energyVad?.minSNR ?? 12,
         snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
+        minEnergy: config?.energyVad?.minEnergy ?? 3e-3
       }
     };
     this.lastSilenceTime = Date.now();

package/dist/index.mjs

@@ -1,13 +1,13 @@
 import "./chunk-WBQAMGXK.mjs";
 import {
   attachProcessingToTrack
-} from "./chunk-Q2I22TJG.mjs";
+} from "./chunk-6F2HZUYO.mjs";
 import {
   createAudioPipeline
-} from "./chunk-SMZJFNRU.mjs";
+} from "./chunk-R5M2DGAQ.mjs";
 import {
   VADStateMachine
-} from "./chunk-DYY2MXMU.mjs";
+} from "./chunk-K4YLH73B.mjs";
 import {
   closeAudioContext,
   getAudioContext,
@@ -21,13 +21,13 @@ import {
   getVADPlugin,
   registerNoiseSuppressionPlugin,
   registerVADPlugin
-} from "./chunk-XZSFQJW4.mjs";
+} from "./chunk-UFKIAMG3.mjs";
 import {
   RNNoisePlugin
 } from "./chunk-XO6B3D4A.mjs";
 import {
   EnergyVADPlugin
-} from "./chunk-KEWK2OKV.mjs";
+} from "./chunk-2G2JFHJY.mjs";
 export {
   EnergyVADPlugin,
   RNNoisePlugin,

package/dist/livekit/integration.js

@@ -127,11 +127,11 @@ 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 ?? 0.05;
-  const minSNR = energyParams.minSNR ?? 10;
+  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 2e-3;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.02;
+  const minSNR = energyParams.minSNR ?? 12;
   const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 1e-3;
+  const minEnergy = energyParams.minEnergy ?? 3e-3;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -160,8 +160,11 @@ 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);
 
@@ -169,36 +172,39 @@ 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 the instantaneous RMS for noise floor tracking to react quickly to silence
-    if (instantRms < this.noiseFloor) {
-      // If signal is quieter than noise floor, adapt downwards quickly
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
+    // 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));
+
+    // Adaptive noise floor estimation using SMOOTHED energy (not instantaneous)
+    // This prevents sharp transients from affecting the noise floor
+    if (this.energy < this.noiseFloor) {
+      // Signal is quieter than noise floor, adapt downwards slowly
+      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + this.energy * this.noiseFloorAdaptRateQuiet;
     } else {
-      // If signal is louder, adapt upwards
-      // We use a multi-stage adaptation rate:
-      // 1. If we are officially speaking, adapt EXTREMELY slowly (0.01x)
-      // 2. If SNR is very high (> 20dB), assume it's speech and adapt very slowly (0.1x)
-      // 3. Otherwise, adapt at the normal loud rate
-      const snr = instantRms / (this.noiseFloor + 1e-6);
-      const snrDb = 20 * Math.log10(Math.max(1e-6, snr));
+      // Calculate SNR based on smoothed energy
+      const smoothedSnr = this.energy / (this.noiseFloor + 1e-6);
+      const smoothedSnrDb = 20 * Math.log10(Math.max(1e-6, smoothedSnr));
       
-      let multiplier = 1.0;
-      if (this.isSpeaking) {
-        multiplier = 0.05;
-      } else if (snrDb > 20) {
-        multiplier = 0.2;
+      // Only adapt upwards if:
+      // 1. SNR is low (< 10dB) - likely just background noise
+      // 2. AND crest factor is low (< 15dB) - not a sharp transient
+      if (smoothedSnrDb < 10 && crestFactorDb < 15) {
+        // This is persistent background noise, adapt upwards
+        this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateLoud) + this.energy * this.noiseFloorAdaptRateLoud;
+      } else {
+        // Either high SNR (speech) or high crest factor (click) - adapt very slowly
+        const slowRate = this.noiseFloorAdaptRateLoud * 0.01;
+        this.noiseFloor = this.noiseFloor * (1 - slowRate) + this.energy * slowRate;
       }
-      
-      const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
-      this.noiseFloor = this.noiseFloor * (1 - adaptRate) + instantRms * adaptRate;
     }
 
     // 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);
+    this.noiseFloor = Math.max(this.noiseFloor, 0.0001);
 
-    // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
+    // SECOND PASS: 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));
     
@@ -207,11 +213,20 @@ 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 });
@@ -334,17 +349,17 @@ 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 ?? 0.01,
-        minSNR: config?.energyVad?.minSNR ?? 10,
+        noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 2e-3,
+        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.02,
+        minSNR: config?.energyVad?.minSNR ?? 12,
         snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
+        minEnergy: config?.energyVad?.minEnergy ?? 3e-3
       }
     };
     this.lastSilenceTime = Date.now();

package/dist/livekit/integration.mjs

@@ -1,12 +1,12 @@
 import {
   attachProcessingToTrack
-} from "../chunk-Q2I22TJG.mjs";
-import "../chunk-SMZJFNRU.mjs";
-import "../chunk-DYY2MXMU.mjs";
+} from "../chunk-6F2HZUYO.mjs";
+import "../chunk-R5M2DGAQ.mjs";
+import "../chunk-K4YLH73B.mjs";
 import "../chunk-OZ7KMC4S.mjs";
-import "../chunk-XZSFQJW4.mjs";
+import "../chunk-UFKIAMG3.mjs";
 import "../chunk-XO6B3D4A.mjs";
-import "../chunk-KEWK2OKV.mjs";
+import "../chunk-2G2JFHJY.mjs";
 export {
   attachProcessingToTrack
 };

package/dist/pipeline/audio-pipeline.js

@@ -125,11 +125,11 @@ 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 ?? 0.05;
-  const minSNR = energyParams.minSNR ?? 10;
+  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 2e-3;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.02;
+  const minSNR = energyParams.minSNR ?? 12;
   const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 1e-3;
+  const minEnergy = energyParams.minEnergy ?? 3e-3;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -158,8 +158,11 @@ 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);
 
@@ -167,36 +170,39 @@ 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 the instantaneous RMS for noise floor tracking to react quickly to silence
-    if (instantRms < this.noiseFloor) {
-      // If signal is quieter than noise floor, adapt downwards quickly
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
+    // 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));
+
+    // Adaptive noise floor estimation using SMOOTHED energy (not instantaneous)
+    // This prevents sharp transients from affecting the noise floor
+    if (this.energy < this.noiseFloor) {
+      // Signal is quieter than noise floor, adapt downwards slowly
+      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + this.energy * this.noiseFloorAdaptRateQuiet;
     } else {
-      // If signal is louder, adapt upwards
-      // We use a multi-stage adaptation rate:
-      // 1. If we are officially speaking, adapt EXTREMELY slowly (0.01x)
-      // 2. If SNR is very high (> 20dB), assume it's speech and adapt very slowly (0.1x)
-      // 3. Otherwise, adapt at the normal loud rate
-      const snr = instantRms / (this.noiseFloor + 1e-6);
-      const snrDb = 20 * Math.log10(Math.max(1e-6, snr));
+      // Calculate SNR based on smoothed energy
+      const smoothedSnr = this.energy / (this.noiseFloor + 1e-6);
+      const smoothedSnrDb = 20 * Math.log10(Math.max(1e-6, smoothedSnr));
       
-      let multiplier = 1.0;
-      if (this.isSpeaking) {
-        multiplier = 0.05;
-      } else if (snrDb > 20) {
-        multiplier = 0.2;
+      // Only adapt upwards if:
+      // 1. SNR is low (< 10dB) - likely just background noise
+      // 2. AND crest factor is low (< 15dB) - not a sharp transient
+      if (smoothedSnrDb < 10 && crestFactorDb < 15) {
+        // This is persistent background noise, adapt upwards
+        this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateLoud) + this.energy * this.noiseFloorAdaptRateLoud;
+      } else {
+        // Either high SNR (speech) or high crest factor (click) - adapt very slowly
+        const slowRate = this.noiseFloorAdaptRateLoud * 0.01;
+        this.noiseFloor = this.noiseFloor * (1 - slowRate) + this.energy * slowRate;
       }
-      
-      const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
-      this.noiseFloor = this.noiseFloor * (1 - adaptRate) + instantRms * adaptRate;
     }
 
     // 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);
+    this.noiseFloor = Math.max(this.noiseFloor, 0.0001);
 
-    // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
+    // SECOND PASS: 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));
     
@@ -205,11 +211,20 @@ 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 });
@@ -332,17 +347,17 @@ 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 ?? 0.01,
-        minSNR: config?.energyVad?.minSNR ?? 10,
+        noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 2e-3,
+        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.02,
+        minSNR: config?.energyVad?.minSNR ?? 12,
         snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
+        minEnergy: config?.energyVad?.minEnergy ?? 3e-3
       }
     };
     this.lastSilenceTime = Date.now();

package/dist/pipeline/audio-pipeline.mjs

@@ -1,11 +1,11 @@
 import {
   createAudioPipeline
-} from "../chunk-SMZJFNRU.mjs";
-import "../chunk-DYY2MXMU.mjs";
+} from "../chunk-R5M2DGAQ.mjs";
+import "../chunk-K4YLH73B.mjs";
 import "../chunk-OZ7KMC4S.mjs";
-import "../chunk-XZSFQJW4.mjs";
+import "../chunk-UFKIAMG3.mjs";
 import "../chunk-XO6B3D4A.mjs";
-import "../chunk-KEWK2OKV.mjs";
+import "../chunk-2G2JFHJY.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;
         /**
@@ -97,17 +97,18 @@ interface AudioProcessingConfig {
             initialNoiseFloor?: number;
             /**
              * Rate at which noise floor adapts to quiet signals (0-1).
-             * Default: 0.01
+             * Default: 0.002 (very slow downward drift)
              */
             noiseFloorAdaptRateQuiet?: number;
             /**
              * Rate at which noise floor adapts to loud signals (0-1).
-             * Default: 0.05 (faster tracking of rising noise)
+             * Applied to low-energy, low-crest-factor signals (background noise).
+             * Default: 0.02
              */
             noiseFloorAdaptRateLoud?: number;
             /**
              * Minimum SNR (Signal-to-Noise Ratio) in dB for speech detection.
-             * Default: 10.0 (more aggressive noise rejection)
+             * Default: 12.0 (aggressive noise rejection)
              */
             minSNR?: number;
             /**
@@ -117,8 +118,8 @@ interface AudioProcessingConfig {
             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 very quiet background noise.
+             * Default: 0.003 (approx -50dB, voice-appropriate 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;
         /**
@@ -97,17 +97,18 @@ interface AudioProcessingConfig {
             initialNoiseFloor?: number;
             /**
              * Rate at which noise floor adapts to quiet signals (0-1).
-             * Default: 0.01
+             * Default: 0.002 (very slow downward drift)
              */
             noiseFloorAdaptRateQuiet?: number;
             /**
              * Rate at which noise floor adapts to loud signals (0-1).
-             * Default: 0.05 (faster tracking of rising noise)
+             * Applied to low-energy, low-crest-factor signals (background noise).
+             * Default: 0.02
              */
             noiseFloorAdaptRateLoud?: number;
             /**
              * Minimum SNR (Signal-to-Noise Ratio) in dB for speech detection.
-             * Default: 10.0 (more aggressive noise rejection)
+             * Default: 12.0 (aggressive noise rejection)
              */
             minSNR?: number;
             /**
@@ -117,8 +118,8 @@ interface AudioProcessingConfig {
             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 very quiet background noise.
+             * Default: 0.003 (approx -50dB, voice-appropriate level)
              */
             minEnergy?: number;
         };

package/dist/vad/vad-node.js

@@ -27,11 +27,11 @@ 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 ?? 0.05;
-  const minSNR = energyParams.minSNR ?? 10;
+  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 2e-3;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.02;
+  const minSNR = energyParams.minSNR ?? 12;
   const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 1e-3;
+  const minEnergy = energyParams.minEnergy ?? 3e-3;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -60,8 +60,11 @@ 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);
 
@@ -69,36 +72,39 @@ 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 the instantaneous RMS for noise floor tracking to react quickly to silence
-    if (instantRms < this.noiseFloor) {
-      // If signal is quieter than noise floor, adapt downwards quickly
-      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
+    // 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));
+
+    // Adaptive noise floor estimation using SMOOTHED energy (not instantaneous)
+    // This prevents sharp transients from affecting the noise floor
+    if (this.energy < this.noiseFloor) {
+      // Signal is quieter than noise floor, adapt downwards slowly
+      this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + this.energy * this.noiseFloorAdaptRateQuiet;
     } else {
-      // If signal is louder, adapt upwards
-      // We use a multi-stage adaptation rate:
-      // 1. If we are officially speaking, adapt EXTREMELY slowly (0.01x)
-      // 2. If SNR is very high (> 20dB), assume it's speech and adapt very slowly (0.1x)
-      // 3. Otherwise, adapt at the normal loud rate
-      const snr = instantRms / (this.noiseFloor + 1e-6);
-      const snrDb = 20 * Math.log10(Math.max(1e-6, snr));
+      // Calculate SNR based on smoothed energy
+      const smoothedSnr = this.energy / (this.noiseFloor + 1e-6);
+      const smoothedSnrDb = 20 * Math.log10(Math.max(1e-6, smoothedSnr));
       
-      let multiplier = 1.0;
-      if (this.isSpeaking) {
-        multiplier = 0.05;
-      } else if (snrDb > 20) {
-        multiplier = 0.2;
+      // Only adapt upwards if:
+      // 1. SNR is low (< 10dB) - likely just background noise
+      // 2. AND crest factor is low (< 15dB) - not a sharp transient
+      if (smoothedSnrDb < 10 && crestFactorDb < 15) {
+        // This is persistent background noise, adapt upwards
+        this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateLoud) + this.energy * this.noiseFloorAdaptRateLoud;
+      } else {
+        // Either high SNR (speech) or high crest factor (click) - adapt very slowly
+        const slowRate = this.noiseFloorAdaptRateLoud * 0.01;
+        this.noiseFloor = this.noiseFloor * (1 - slowRate) + this.energy * slowRate;
       }
-      
-      const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
-      this.noiseFloor = this.noiseFloor * (1 - adaptRate) + instantRms * adaptRate;
     }
 
     // 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);
+    this.noiseFloor = Math.max(this.noiseFloor, 0.0001);
 
-    // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
+    // SECOND PASS: 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));
     
@@ -107,11 +113,20 @@ 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-KEWK2OKV.mjs";
+} from "../chunk-2G2JFHJY.mjs";
 export {
   EnergyVADPlugin
 };

package/dist/vad/vad-state.js

@@ -44,17 +44,17 @@ 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 ?? 0.01,
-        minSNR: config?.energyVad?.minSNR ?? 10,
+        noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 2e-3,
+        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.02,
+        minSNR: config?.energyVad?.minSNR ?? 12,
         snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
+        minEnergy: config?.energyVad?.minEnergy ?? 3e-3
       }
     };
     this.lastSilenceTime = Date.now();

package/dist/vad/vad-state.mjs

@@ -1,6 +1,6 @@
 import {
   VADStateMachine
-} from "../chunk-DYY2MXMU.mjs";
+} from "../chunk-K4YLH73B.mjs";
 export {
   VADStateMachine
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tensamin/audio",
-  "version": "0.1.12",
+  "version": "0.1.14",
   "main": "dist/index.js",
   "module": "dist/index.mjs",
   "types": "dist/index.d.ts",