@tensamin/audio 0.1.10 → 0.1.12

package/README.md

@@ -101,11 +101,11 @@ vad: {
   energyVad?: {
     smoothing: number;                 // Default: 0.95
     initialNoiseFloor: number;         // Default: 0.001
-    noiseFloorAdaptRateQuiet: number;  // Default: 0.005
-    noiseFloorAdaptRateLoud: number;   // Default: 0.01
+    noiseFloorAdaptRateQuiet: number;  // Default: 0.01
+    noiseFloorAdaptRateLoud: number;   // Default: 0.05
     minSNR: number;                    // Default: 10.0 (dB)
     snrRange: number;                  // Default: 10.0 (dB)
-    minEnergy: number;                 // Default: 0.0005
+    minEnergy: number;                 // Default: 0.001
   };
 }
 ```
@@ -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
+- `minSpeechDurationMs`: Minimum duration to consider as valid speech (Default: 150ms)
 - `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
+- `minEnergy`: Minimum absolute RMS energy to consider as speech (Default: 0.001, ~-60dB)
 
 ### Output Control
 

package/dist/{chunk-VCQMZVO3.mjs → chunk-DYY2MXMU.mjs}

@@ -30,7 +30,7 @@ var VADStateMachine = class {
         noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.01,
         minSNR: config?.energyVad?.minSNR ?? 10,
         snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 5e-4
+        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
       }
     };
     this.lastSilenceTime = Date.now();
@@ -61,25 +61,33 @@ var VADStateMachine = class {
         newState = "silent";
         this.lastSilenceTime = timestamp;
       }
-    } else if (this.currentState === "speech_starting" || this.currentState === "speaking") {
+    } else if (this.currentState === "speech_starting") {
+      if (probability >= stopThreshold) {
+        const speechDuration = timestamp - this.speechStartTime;
+        if (speechDuration >= minSpeechDurationMs) {
+          newState = "speaking";
+        } else {
+          newState = "speech_starting";
+        }
+        this.lastSpeechTime = timestamp;
+      } else {
+        newState = "silent";
+        this.lastSilenceTime = timestamp;
+      }
+    } else if (this.currentState === "speaking") {
       if (probability >= stopThreshold) {
         newState = "speaking";
         this.lastSpeechTime = timestamp;
       } else {
         const timeSinceSpeech = timestamp - this.lastSpeechTime;
-        const speechDuration = timestamp - this.speechStartTime;
         if (timeSinceSpeech < hangoverMs) {
           newState = "speaking";
-        } else if (speechDuration < minSpeechDurationMs) {
-          newState = "silent";
-          this.lastSilenceTime = timestamp;
         } else {
           newState = "speech_ending";
           this.lastSilenceTime = timestamp;
         }
       }
     }
-    if (newState === "speech_starting") newState = "speaking";
     if (newState === "speech_ending") newState = "silent";
     this.currentState = newState;
     return {

package/dist/{chunk-SMNOCQYR.mjs → chunk-KEWK2OKV.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 ?? 5e-3;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.01;
+  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.01;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.05;
   const minSNR = energyParams.minSNR ?? 10;
   const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 5e-4;
+  const minEnergy = energyParams.minEnergy ?? 1e-3;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -61,9 +61,9 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
       
       let multiplier = 1.0;
       if (this.isSpeaking) {
-        multiplier = 0.01;
+        multiplier = 0.05;
       } else if (snrDb > 20) {
-        multiplier = 0.1;
+        multiplier = 0.2;
       }
       
       const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
@@ -71,8 +71,8 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     }
 
     // Ensure noise floor doesn't drop to absolute zero
-    // 0.0002 is approx -74dB, a reasonable floor for ambient noise
-    this.noiseFloor = Math.max(this.noiseFloor, 0.0002);
+    // 0.00005 is approx -86dB, very quiet but prevents SNR explosion
+    this.noiseFloor = Math.max(this.noiseFloor, 0.00005);
 
     // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
     const snr = this.energy / (this.noiseFloor + 1e-6);
@@ -84,8 +84,10 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     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
     if (this.energy < this.minEnergy) {
-      probability = 0;
+      const energyRatio = this.energy / (this.minEnergy + 1e-6);
+      probability *= Math.pow(energyRatio, 2); // Quadratic falloff
     }
 
     this.port.postMessage({ probability, snr: snrDb, noiseFloor: this.noiseFloor, rms: this.energy });

package/dist/{chunk-IL4F7WVW.mjs → chunk-Q2I22TJG.mjs}

@@ -1,6 +1,6 @@
 import {
   createAudioPipeline
-} from "./chunk-CD5XFC5M.mjs";
+} from "./chunk-SMZJFNRU.mjs";
 
 // src/livekit/integration.ts
 async function attachProcessingToTrack(track, config = {}) {

package/dist/{chunk-CD5XFC5M.mjs → chunk-SMZJFNRU.mjs}

@@ -1,6 +1,6 @@
 import {
   VADStateMachine
-} from "./chunk-VCQMZVO3.mjs";
+} from "./chunk-DYY2MXMU.mjs";
 import {
   getAudioContext,
   registerPipeline,
@@ -9,7 +9,7 @@ import {
 import {
   getNoiseSuppressionPlugin,
   getVADPlugin
-} from "./chunk-Z3QBDLTM.mjs";
+} from "./chunk-XZSFQJW4.mjs";
 
 // src/pipeline/audio-pipeline.ts
 import mitt from "mitt";

package/dist/{chunk-Z3QBDLTM.mjs → chunk-XZSFQJW4.mjs}

@@ -3,7 +3,7 @@ import {
 } from "./chunk-XO6B3D4A.mjs";
 import {
   EnergyVADPlugin
-} from "./chunk-SMNOCQYR.mjs";
+} from "./chunk-KEWK2OKV.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 ?? 5e-3;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.01;
+  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.01;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.05;
   const minSNR = energyParams.minSNR ?? 10;
   const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 5e-4;
+  const minEnergy = energyParams.minEnergy ?? 1e-3;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -164,9 +164,9 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
       
       let multiplier = 1.0;
       if (this.isSpeaking) {
-        multiplier = 0.01;
+        multiplier = 0.05;
       } else if (snrDb > 20) {
-        multiplier = 0.1;
+        multiplier = 0.2;
       }
       
       const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
@@ -174,8 +174,8 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     }
 
     // Ensure noise floor doesn't drop to absolute zero
-    // 0.0002 is approx -74dB, a reasonable floor for ambient noise
-    this.noiseFloor = Math.max(this.noiseFloor, 0.0002);
+    // 0.00005 is approx -86dB, very quiet but prevents SNR explosion
+    this.noiseFloor = Math.max(this.noiseFloor, 0.00005);
 
     // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
     const snr = this.energy / (this.noiseFloor + 1e-6);
@@ -187,8 +187,10 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     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
     if (this.energy < this.minEnergy) {
-      probability = 0;
+      const energyRatio = this.energy / (this.minEnergy + 1e-6);
+      probability *= Math.pow(energyRatio, 2); // Quadratic falloff
     }
 
     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-Z3QBDLTM.mjs";
+} from "../chunk-XZSFQJW4.mjs";
 import "../chunk-XO6B3D4A.mjs";
-import "../chunk-SMNOCQYR.mjs";
+import "../chunk-KEWK2OKV.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 ?? 5e-3;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.01;
+  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.01;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.05;
   const minSNR = energyParams.minSNR ?? 10;
   const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 5e-4;
+  const minEnergy = energyParams.minEnergy ?? 1e-3;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -216,9 +216,9 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
       
       let multiplier = 1.0;
       if (this.isSpeaking) {
-        multiplier = 0.01;
+        multiplier = 0.05;
       } else if (snrDb > 20) {
-        multiplier = 0.1;
+        multiplier = 0.2;
       }
       
       const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
@@ -226,8 +226,8 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     }
 
     // Ensure noise floor doesn't drop to absolute zero
-    // 0.0002 is approx -74dB, a reasonable floor for ambient noise
-    this.noiseFloor = Math.max(this.noiseFloor, 0.0002);
+    // 0.00005 is approx -86dB, very quiet but prevents SNR explosion
+    this.noiseFloor = Math.max(this.noiseFloor, 0.00005);
 
     // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
     const snr = this.energy / (this.noiseFloor + 1e-6);
@@ -239,8 +239,10 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     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
     if (this.energy < this.minEnergy) {
-      probability = 0;
+      const energyRatio = this.energy / (this.minEnergy + 1e-6);
+      probability *= Math.pow(energyRatio, 2); // Quadratic falloff
     }
 
     this.port.postMessage({ probability, snr: snrDb, noiseFloor: this.noiseFloor, rms: this.energy });
@@ -379,7 +381,7 @@ var VADStateMachine = class {
         noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.01,
         minSNR: config?.energyVad?.minSNR ?? 10,
         snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 5e-4
+        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
       }
     };
     this.lastSilenceTime = Date.now();
@@ -410,25 +412,33 @@ var VADStateMachine = class {
         newState = "silent";
         this.lastSilenceTime = timestamp;
       }
-    } else if (this.currentState === "speech_starting" || this.currentState === "speaking") {
+    } else if (this.currentState === "speech_starting") {
+      if (probability >= stopThreshold) {
+        const speechDuration = timestamp - this.speechStartTime;
+        if (speechDuration >= minSpeechDurationMs) {
+          newState = "speaking";
+        } else {
+          newState = "speech_starting";
+        }
+        this.lastSpeechTime = timestamp;
+      } else {
+        newState = "silent";
+        this.lastSilenceTime = timestamp;
+      }
+    } else if (this.currentState === "speaking") {
       if (probability >= stopThreshold) {
         newState = "speaking";
         this.lastSpeechTime = timestamp;
       } else {
         const timeSinceSpeech = timestamp - this.lastSpeechTime;
-        const speechDuration = timestamp - this.speechStartTime;
         if (timeSinceSpeech < hangoverMs) {
           newState = "speaking";
-        } else if (speechDuration < minSpeechDurationMs) {
-          newState = "silent";
-          this.lastSilenceTime = timestamp;
         } else {
           newState = "speech_ending";
           this.lastSilenceTime = timestamp;
         }
       }
     }
-    if (newState === "speech_starting") newState = "speaking";
     if (newState === "speech_ending") newState = "silent";
     this.currentState = newState;
     return {

package/dist/index.mjs

@@ -1,13 +1,13 @@
 import "./chunk-WBQAMGXK.mjs";
 import {
   attachProcessingToTrack
-} from "./chunk-IL4F7WVW.mjs";
+} from "./chunk-Q2I22TJG.mjs";
 import {
   createAudioPipeline
-} from "./chunk-CD5XFC5M.mjs";
+} from "./chunk-SMZJFNRU.mjs";
 import {
   VADStateMachine
-} from "./chunk-VCQMZVO3.mjs";
+} from "./chunk-DYY2MXMU.mjs";
 import {
   closeAudioContext,
   getAudioContext,
@@ -21,13 +21,13 @@ import {
   getVADPlugin,
   registerNoiseSuppressionPlugin,
   registerVADPlugin
-} from "./chunk-Z3QBDLTM.mjs";
+} from "./chunk-XZSFQJW4.mjs";
 import {
   RNNoisePlugin
 } from "./chunk-XO6B3D4A.mjs";
 import {
   EnergyVADPlugin
-} from "./chunk-SMNOCQYR.mjs";
+} from "./chunk-KEWK2OKV.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 ?? 5e-3;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.01;
+  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.01;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.05;
   const minSNR = energyParams.minSNR ?? 10;
   const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 5e-4;
+  const minEnergy = energyParams.minEnergy ?? 1e-3;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -185,9 +185,9 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
       
       let multiplier = 1.0;
       if (this.isSpeaking) {
-        multiplier = 0.01;
+        multiplier = 0.05;
       } else if (snrDb > 20) {
-        multiplier = 0.1;
+        multiplier = 0.2;
       }
       
       const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
@@ -195,8 +195,8 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     }
 
     // Ensure noise floor doesn't drop to absolute zero
-    // 0.0002 is approx -74dB, a reasonable floor for ambient noise
-    this.noiseFloor = Math.max(this.noiseFloor, 0.0002);
+    // 0.00005 is approx -86dB, very quiet but prevents SNR explosion
+    this.noiseFloor = Math.max(this.noiseFloor, 0.00005);
 
     // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
     const snr = this.energy / (this.noiseFloor + 1e-6);
@@ -208,8 +208,10 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     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
     if (this.energy < this.minEnergy) {
-      probability = 0;
+      const energyRatio = this.energy / (this.minEnergy + 1e-6);
+      probability *= Math.pow(energyRatio, 2); // Quadratic falloff
     }
 
     this.port.postMessage({ probability, snr: snrDb, noiseFloor: this.noiseFloor, rms: this.energy });
@@ -342,7 +344,7 @@ var VADStateMachine = class {
         noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.01,
         minSNR: config?.energyVad?.minSNR ?? 10,
         snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 5e-4
+        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
       }
     };
     this.lastSilenceTime = Date.now();
@@ -373,25 +375,33 @@ var VADStateMachine = class {
         newState = "silent";
         this.lastSilenceTime = timestamp;
       }
-    } else if (this.currentState === "speech_starting" || this.currentState === "speaking") {
+    } else if (this.currentState === "speech_starting") {
+      if (probability >= stopThreshold) {
+        const speechDuration = timestamp - this.speechStartTime;
+        if (speechDuration >= minSpeechDurationMs) {
+          newState = "speaking";
+        } else {
+          newState = "speech_starting";
+        }
+        this.lastSpeechTime = timestamp;
+      } else {
+        newState = "silent";
+        this.lastSilenceTime = timestamp;
+      }
+    } else if (this.currentState === "speaking") {
       if (probability >= stopThreshold) {
         newState = "speaking";
         this.lastSpeechTime = timestamp;
       } else {
         const timeSinceSpeech = timestamp - this.lastSpeechTime;
-        const speechDuration = timestamp - this.speechStartTime;
         if (timeSinceSpeech < hangoverMs) {
           newState = "speaking";
-        } else if (speechDuration < minSpeechDurationMs) {
-          newState = "silent";
-          this.lastSilenceTime = timestamp;
         } else {
           newState = "speech_ending";
           this.lastSilenceTime = timestamp;
         }
       }
     }
-    if (newState === "speech_starting") newState = "speaking";
     if (newState === "speech_ending") newState = "silent";
     this.currentState = newState;
     return {

package/dist/livekit/integration.mjs

@@ -1,12 +1,12 @@
 import {
   attachProcessingToTrack
-} from "../chunk-IL4F7WVW.mjs";
-import "../chunk-CD5XFC5M.mjs";
-import "../chunk-VCQMZVO3.mjs";
+} from "../chunk-Q2I22TJG.mjs";
+import "../chunk-SMZJFNRU.mjs";
+import "../chunk-DYY2MXMU.mjs";
 import "../chunk-OZ7KMC4S.mjs";
-import "../chunk-Z3QBDLTM.mjs";
+import "../chunk-XZSFQJW4.mjs";
 import "../chunk-XO6B3D4A.mjs";
-import "../chunk-SMNOCQYR.mjs";
+import "../chunk-KEWK2OKV.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 ?? 5e-3;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.01;
+  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.01;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.05;
   const minSNR = energyParams.minSNR ?? 10;
   const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 5e-4;
+  const minEnergy = energyParams.minEnergy ?? 1e-3;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -183,9 +183,9 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
       
       let multiplier = 1.0;
       if (this.isSpeaking) {
-        multiplier = 0.01;
+        multiplier = 0.05;
       } else if (snrDb > 20) {
-        multiplier = 0.1;
+        multiplier = 0.2;
       }
       
       const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
@@ -193,8 +193,8 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     }
 
     // Ensure noise floor doesn't drop to absolute zero
-    // 0.0002 is approx -74dB, a reasonable floor for ambient noise
-    this.noiseFloor = Math.max(this.noiseFloor, 0.0002);
+    // 0.00005 is approx -86dB, very quiet but prevents SNR explosion
+    this.noiseFloor = Math.max(this.noiseFloor, 0.00005);
 
     // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
     const snr = this.energy / (this.noiseFloor + 1e-6);
@@ -206,8 +206,10 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     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
     if (this.energy < this.minEnergy) {
-      probability = 0;
+      const energyRatio = this.energy / (this.minEnergy + 1e-6);
+      probability *= Math.pow(energyRatio, 2); // Quadratic falloff
     }
 
     this.port.postMessage({ probability, snr: snrDb, noiseFloor: this.noiseFloor, rms: this.energy });
@@ -340,7 +342,7 @@ var VADStateMachine = class {
         noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.01,
         minSNR: config?.energyVad?.minSNR ?? 10,
         snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 5e-4
+        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
       }
     };
     this.lastSilenceTime = Date.now();
@@ -371,25 +373,33 @@ var VADStateMachine = class {
         newState = "silent";
         this.lastSilenceTime = timestamp;
       }
-    } else if (this.currentState === "speech_starting" || this.currentState === "speaking") {
+    } else if (this.currentState === "speech_starting") {
+      if (probability >= stopThreshold) {
+        const speechDuration = timestamp - this.speechStartTime;
+        if (speechDuration >= minSpeechDurationMs) {
+          newState = "speaking";
+        } else {
+          newState = "speech_starting";
+        }
+        this.lastSpeechTime = timestamp;
+      } else {
+        newState = "silent";
+        this.lastSilenceTime = timestamp;
+      }
+    } else if (this.currentState === "speaking") {
       if (probability >= stopThreshold) {
         newState = "speaking";
         this.lastSpeechTime = timestamp;
       } else {
         const timeSinceSpeech = timestamp - this.lastSpeechTime;
-        const speechDuration = timestamp - this.speechStartTime;
         if (timeSinceSpeech < hangoverMs) {
           newState = "speaking";
-        } else if (speechDuration < minSpeechDurationMs) {
-          newState = "silent";
-          this.lastSilenceTime = timestamp;
         } else {
           newState = "speech_ending";
           this.lastSilenceTime = timestamp;
         }
       }
     }
-    if (newState === "speech_starting") newState = "speaking";
     if (newState === "speech_ending") newState = "silent";
     this.currentState = newState;
     return {

package/dist/pipeline/audio-pipeline.mjs

@@ -1,11 +1,11 @@
 import {
   createAudioPipeline
-} from "../chunk-CD5XFC5M.mjs";
-import "../chunk-VCQMZVO3.mjs";
+} from "../chunk-SMZJFNRU.mjs";
+import "../chunk-DYY2MXMU.mjs";
 import "../chunk-OZ7KMC4S.mjs";
-import "../chunk-Z3QBDLTM.mjs";
+import "../chunk-XZSFQJW4.mjs";
 import "../chunk-XO6B3D4A.mjs";
-import "../chunk-SMNOCQYR.mjs";
+import "../chunk-KEWK2OKV.mjs";
 export {
   createAudioPipeline
 };

package/dist/types.d.mts

@@ -97,12 +97,12 @@ interface AudioProcessingConfig {
             initialNoiseFloor?: number;
             /**
              * Rate at which noise floor adapts to quiet signals (0-1).
-             * Default: 0.005 (slower downward drift)
+             * Default: 0.01
              */
             noiseFloorAdaptRateQuiet?: number;
             /**
              * Rate at which noise floor adapts to loud signals (0-1).
-             * Default: 0.01
+             * Default: 0.05 (faster tracking of rising noise)
              */
             noiseFloorAdaptRateLoud?: number;
             /**
@@ -118,7 +118,7 @@ interface AudioProcessingConfig {
             /**
              * Minimum absolute RMS energy to consider as speech.
              * Prevents triggering on very quiet background noise in silent rooms.
-             * Default: 0.0005
+             * Default: 0.001 (approx -60dB)
              */
             minEnergy?: number;
         };

package/dist/types.d.ts

@@ -97,12 +97,12 @@ interface AudioProcessingConfig {
             initialNoiseFloor?: number;
             /**
              * Rate at which noise floor adapts to quiet signals (0-1).
-             * Default: 0.005 (slower downward drift)
+             * Default: 0.01
              */
             noiseFloorAdaptRateQuiet?: number;
             /**
              * Rate at which noise floor adapts to loud signals (0-1).
-             * Default: 0.01
+             * Default: 0.05 (faster tracking of rising noise)
              */
             noiseFloorAdaptRateLoud?: number;
             /**
@@ -118,7 +118,7 @@ interface AudioProcessingConfig {
             /**
              * Minimum absolute RMS energy to consider as speech.
              * Prevents triggering on very quiet background noise in silent rooms.
-             * Default: 0.0005
+             * Default: 0.001 (approx -60dB)
              */
             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 ?? 5e-3;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.01;
+  const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.01;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.05;
   const minSNR = energyParams.minSNR ?? 10;
   const snrRange = energyParams.snrRange ?? 10;
-  const minEnergy = energyParams.minEnergy ?? 5e-4;
+  const minEnergy = energyParams.minEnergy ?? 1e-3;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
   constructor() {
@@ -85,9 +85,9 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
       
       let multiplier = 1.0;
       if (this.isSpeaking) {
-        multiplier = 0.01;
+        multiplier = 0.05;
       } else if (snrDb > 20) {
-        multiplier = 0.1;
+        multiplier = 0.2;
       }
       
       const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
@@ -95,8 +95,8 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     }
 
     // Ensure noise floor doesn't drop to absolute zero
-    // 0.0002 is approx -74dB, a reasonable floor for ambient noise
-    this.noiseFloor = Math.max(this.noiseFloor, 0.0002);
+    // 0.00005 is approx -86dB, very quiet but prevents SNR explosion
+    this.noiseFloor = Math.max(this.noiseFloor, 0.00005);
 
     // Calculate Signal-to-Noise Ratio (SNR) in dB using smoothed energy
     const snr = this.energy / (this.noiseFloor + 1e-6);
@@ -108,8 +108,10 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
     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
     if (this.energy < this.minEnergy) {
-      probability = 0;
+      const energyRatio = this.energy / (this.minEnergy + 1e-6);
+      probability *= Math.pow(energyRatio, 2); // Quadratic falloff
     }
 
     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-SMNOCQYR.mjs";
+} from "../chunk-KEWK2OKV.mjs";
 export {
   EnergyVADPlugin
 };

package/dist/vad/vad-state.js

@@ -54,7 +54,7 @@ var VADStateMachine = class {
         noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.01,
         minSNR: config?.energyVad?.minSNR ?? 10,
         snrRange: config?.energyVad?.snrRange ?? 10,
-        minEnergy: config?.energyVad?.minEnergy ?? 5e-4
+        minEnergy: config?.energyVad?.minEnergy ?? 1e-3
       }
     };
     this.lastSilenceTime = Date.now();
@@ -85,25 +85,33 @@ var VADStateMachine = class {
         newState = "silent";
         this.lastSilenceTime = timestamp;
       }
-    } else if (this.currentState === "speech_starting" || this.currentState === "speaking") {
+    } else if (this.currentState === "speech_starting") {
+      if (probability >= stopThreshold) {
+        const speechDuration = timestamp - this.speechStartTime;
+        if (speechDuration >= minSpeechDurationMs) {
+          newState = "speaking";
+        } else {
+          newState = "speech_starting";
+        }
+        this.lastSpeechTime = timestamp;
+      } else {
+        newState = "silent";
+        this.lastSilenceTime = timestamp;
+      }
+    } else if (this.currentState === "speaking") {
       if (probability >= stopThreshold) {
         newState = "speaking";
         this.lastSpeechTime = timestamp;
       } else {
         const timeSinceSpeech = timestamp - this.lastSpeechTime;
-        const speechDuration = timestamp - this.speechStartTime;
         if (timeSinceSpeech < hangoverMs) {
           newState = "speaking";
-        } else if (speechDuration < minSpeechDurationMs) {
-          newState = "silent";
-          this.lastSilenceTime = timestamp;
         } else {
           newState = "speech_ending";
           this.lastSilenceTime = timestamp;
         }
       }
     }
-    if (newState === "speech_starting") newState = "speaking";
     if (newState === "speech_ending") newState = "silent";
     this.currentState = newState;
     return {

package/dist/vad/vad-state.mjs

@@ -1,6 +1,6 @@
 import {
   VADStateMachine
-} from "../chunk-VCQMZVO3.mjs";
+} from "../chunk-DYY2MXMU.mjs";
 export {
   VADStateMachine
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tensamin/audio",
-  "version": "0.1.10",
+  "version": "0.1.12",
   "main": "dist/index.js",
   "module": "dist/index.mjs",
   "types": "dist/index.d.ts",