@tensamin/audio 0.1.7 → 0.1.8

package/README.md

@@ -102,9 +102,9 @@ vad: {
     smoothing: number;                 // Default: 0.95
     initialNoiseFloor: number;         // Default: 0.001
     noiseFloorAdaptRateQuiet: number;  // Default: 0.05
-    noiseFloorAdaptRateLoud: number;   // Default: 0.005
-    minSNR: number;                    // Default: 6.0 (dB)
-    snrRange: number;                  // Default: 12.0 (dB)
+    noiseFloorAdaptRateLoud: number;   // Default: 0.01
+    minSNR: number;                    // Default: 10.0 (dB)
+    snrRange: number;                  // Default: 10.0 (dB)
     minEnergy: number;                 // Default: 0.0005
   };
 }
@@ -112,9 +112,9 @@ vad: {
 
 **Threshold Parameters:**
 
-- `startThreshold`: Probability threshold to unmute audio (Default: 0.6, ~13.2dB SNR)
-- `stopThreshold`: Probability threshold to mute audio (Default: 0.45, ~11.4dB SNR)
-- `hangoverMs`: Delay before muting after speech stops
+- `startThreshold`: Probability threshold to unmute audio (Default: 0.8, ~18dB SNR)
+- `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
 - `minSilenceDurationMs`: Minimum silence duration between speech segments

package/dist/{chunk-JP6DA62Y.mjs → chunk-2UPI6VWY.mjs}

@@ -1,6 +1,6 @@
 import {
   VADStateMachine
-} from "./chunk-2EX3FXSF.mjs";
+} from "./chunk-XHMNP7NC.mjs";
 import {
   getAudioContext,
   registerPipeline,
@@ -9,7 +9,7 @@ import {
 import {
   getNoiseSuppressionPlugin,
   getVADPlugin
-} from "./chunk-BMVZ3KKG.mjs";
+} from "./chunk-FOGC2MFA.mjs";
 
 // src/pipeline/audio-pipeline.ts
 import mitt from "mitt";

package/dist/{chunk-2TKYGFMC.mjs → chunk-3A2CTC4K.mjs}

@@ -4,9 +4,9 @@ var createEnergyVadWorkletCode = (vadConfig) => {
   const smoothing = energyParams.smoothing ?? 0.95;
   const initialNoiseFloor = energyParams.initialNoiseFloor ?? 1e-3;
   const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.05;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 5e-3;
-  const minSNR = energyParams.minSNR ?? 6;
-  const snrRange = energyParams.snrRange ?? 12;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.01;
+  const minSNR = energyParams.minSNR ?? 10;
+  const snrRange = energyParams.snrRange ?? 10;
   const minEnergy = energyParams.minEnergy ?? 5e-4;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
@@ -52,9 +52,21 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
       this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
     } else {
       // If signal is louder, adapt upwards
-      // If we are currently speaking, adapt EXTREMELY slowly to avoid "chasing" speech
-      // If we are silent, adapt at the normal loud rate
-      const adaptRate = this.isSpeaking ? (this.noiseFloorAdaptRateLoud * 0.02) : this.noiseFloorAdaptRateLoud;
+      // 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));
+      
+      let multiplier = 1.0;
+      if (this.isSpeaking) {
+        multiplier = 0.01;
+      } else if (snrDb > 20) {
+        multiplier = 0.1;
+      }
+      
+      const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
       this.noiseFloor = this.noiseFloor * (1 - adaptRate) + instantRms * adaptRate;
     }
 

package/dist/{chunk-BMVZ3KKG.mjs → chunk-FOGC2MFA.mjs}

@@ -3,7 +3,7 @@ import {
 } from "./chunk-XO6B3D4A.mjs";
 import {
   EnergyVADPlugin
-} from "./chunk-2TKYGFMC.mjs";
+} from "./chunk-3A2CTC4K.mjs";
 
 // src/extensibility/plugins.ts
 var nsPlugins = /* @__PURE__ */ new Map();

package/dist/{chunk-2EX3FXSF.mjs → chunk-XHMNP7NC.mjs}

@@ -12,11 +12,11 @@ var VADStateMachine = class {
       enabled: config?.enabled ?? true,
       pluginName: config?.pluginName ?? "energy-vad",
       // Voice-optimized defaults
-      startThreshold: config?.startThreshold ?? 0.6,
+      startThreshold: config?.startThreshold ?? 0.8,
       // Higher threshold to avoid noise
-      stopThreshold: config?.stopThreshold ?? 0.45,
+      stopThreshold: config?.stopThreshold ?? 0.3,
       // Balanced for voice
-      hangoverMs: config?.hangoverMs ?? 400,
+      hangoverMs: config?.hangoverMs ?? 300,
       // Smooth for natural speech
       preRollMs: config?.preRollMs ?? 250,
       // Generous pre-roll
@@ -26,9 +26,9 @@ var VADStateMachine = class {
         smoothing: config?.energyVad?.smoothing ?? 0.95,
         initialNoiseFloor: config?.energyVad?.initialNoiseFloor ?? 1e-3,
         noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 0.05,
-        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 5e-3,
-        minSNR: config?.energyVad?.minSNR ?? 6,
-        snrRange: config?.energyVad?.snrRange ?? 12,
+        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.01,
+        minSNR: config?.energyVad?.minSNR ?? 10,
+        snrRange: config?.energyVad?.snrRange ?? 10,
         minEnergy: config?.energyVad?.minEnergy ?? 5e-4
       }
     };

package/dist/{chunk-UQG6Z5W3.mjs → chunk-Y6IG7XGC.mjs}

@@ -1,6 +1,6 @@
 import {
   createAudioPipeline
-} from "./chunk-JP6DA62Y.mjs";
+} from "./chunk-2UPI6VWY.mjs";
 
 // src/livekit/integration.ts
 async function attachProcessingToTrack(track, config = {}) {

package/dist/extensibility/plugins.js

@@ -107,9 +107,9 @@ var createEnergyVadWorkletCode = (vadConfig) => {
   const smoothing = energyParams.smoothing ?? 0.95;
   const initialNoiseFloor = energyParams.initialNoiseFloor ?? 1e-3;
   const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.05;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 5e-3;
-  const minSNR = energyParams.minSNR ?? 6;
-  const snrRange = energyParams.snrRange ?? 12;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.01;
+  const minSNR = energyParams.minSNR ?? 10;
+  const snrRange = energyParams.snrRange ?? 10;
   const minEnergy = energyParams.minEnergy ?? 5e-4;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
@@ -155,9 +155,21 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
       this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
     } else {
       // If signal is louder, adapt upwards
-      // If we are currently speaking, adapt EXTREMELY slowly to avoid "chasing" speech
-      // If we are silent, adapt at the normal loud rate
-      const adaptRate = this.isSpeaking ? (this.noiseFloorAdaptRateLoud * 0.02) : this.noiseFloorAdaptRateLoud;
+      // 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));
+      
+      let multiplier = 1.0;
+      if (this.isSpeaking) {
+        multiplier = 0.01;
+      } else if (snrDb > 20) {
+        multiplier = 0.1;
+      }
+      
+      const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
       this.noiseFloor = this.noiseFloor * (1 - adaptRate) + instantRms * adaptRate;
     }
 

package/dist/extensibility/plugins.mjs

@@ -3,9 +3,9 @@ import {
   getVADPlugin,
   registerNoiseSuppressionPlugin,
   registerVADPlugin
-} from "../chunk-BMVZ3KKG.mjs";
+} from "../chunk-FOGC2MFA.mjs";
 import "../chunk-XO6B3D4A.mjs";
-import "../chunk-2TKYGFMC.mjs";
+import "../chunk-3A2CTC4K.mjs";
 export {
   getNoiseSuppressionPlugin,
   getVADPlugin,

package/dist/index.js

@@ -159,9 +159,9 @@ var createEnergyVadWorkletCode = (vadConfig) => {
   const smoothing = energyParams.smoothing ?? 0.95;
   const initialNoiseFloor = energyParams.initialNoiseFloor ?? 1e-3;
   const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.05;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 5e-3;
-  const minSNR = energyParams.minSNR ?? 6;
-  const snrRange = energyParams.snrRange ?? 12;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.01;
+  const minSNR = energyParams.minSNR ?? 10;
+  const snrRange = energyParams.snrRange ?? 10;
   const minEnergy = energyParams.minEnergy ?? 5e-4;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
@@ -207,9 +207,21 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
       this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
     } else {
       // If signal is louder, adapt upwards
-      // If we are currently speaking, adapt EXTREMELY slowly to avoid "chasing" speech
-      // If we are silent, adapt at the normal loud rate
-      const adaptRate = this.isSpeaking ? (this.noiseFloorAdaptRateLoud * 0.02) : this.noiseFloorAdaptRateLoud;
+      // 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));
+      
+      let multiplier = 1.0;
+      if (this.isSpeaking) {
+        multiplier = 0.01;
+      } else if (snrDb > 20) {
+        multiplier = 0.1;
+      }
+      
+      const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
       this.noiseFloor = this.noiseFloor * (1 - adaptRate) + instantRms * adaptRate;
     }
 
@@ -348,11 +360,11 @@ var VADStateMachine = class {
       enabled: config?.enabled ?? true,
       pluginName: config?.pluginName ?? "energy-vad",
       // Voice-optimized defaults
-      startThreshold: config?.startThreshold ?? 0.6,
+      startThreshold: config?.startThreshold ?? 0.8,
       // Higher threshold to avoid noise
-      stopThreshold: config?.stopThreshold ?? 0.45,
+      stopThreshold: config?.stopThreshold ?? 0.3,
       // Balanced for voice
-      hangoverMs: config?.hangoverMs ?? 400,
+      hangoverMs: config?.hangoverMs ?? 300,
       // Smooth for natural speech
       preRollMs: config?.preRollMs ?? 250,
       // Generous pre-roll
@@ -362,9 +374,9 @@ var VADStateMachine = class {
         smoothing: config?.energyVad?.smoothing ?? 0.95,
         initialNoiseFloor: config?.energyVad?.initialNoiseFloor ?? 1e-3,
         noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 0.05,
-        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 5e-3,
-        minSNR: config?.energyVad?.minSNR ?? 6,
-        snrRange: config?.energyVad?.snrRange ?? 12,
+        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.01,
+        minSNR: config?.energyVad?.minSNR ?? 10,
+        snrRange: config?.energyVad?.snrRange ?? 10,
         minEnergy: config?.energyVad?.minEnergy ?? 5e-4
       }
     };

package/dist/index.mjs

@@ -1,13 +1,13 @@
 import "./chunk-WBQAMGXK.mjs";
 import {
   attachProcessingToTrack
-} from "./chunk-UQG6Z5W3.mjs";
+} from "./chunk-Y6IG7XGC.mjs";
 import {
   createAudioPipeline
-} from "./chunk-JP6DA62Y.mjs";
+} from "./chunk-2UPI6VWY.mjs";
 import {
   VADStateMachine
-} from "./chunk-2EX3FXSF.mjs";
+} from "./chunk-XHMNP7NC.mjs";
 import {
   closeAudioContext,
   getAudioContext,
@@ -21,13 +21,13 @@ import {
   getVADPlugin,
   registerNoiseSuppressionPlugin,
   registerVADPlugin
-} from "./chunk-BMVZ3KKG.mjs";
+} from "./chunk-FOGC2MFA.mjs";
 import {
   RNNoisePlugin
 } from "./chunk-XO6B3D4A.mjs";
 import {
   EnergyVADPlugin
-} from "./chunk-2TKYGFMC.mjs";
+} from "./chunk-3A2CTC4K.mjs";
 export {
   EnergyVADPlugin,
   RNNoisePlugin,

package/dist/livekit/integration.js

@@ -128,9 +128,9 @@ var createEnergyVadWorkletCode = (vadConfig) => {
   const smoothing = energyParams.smoothing ?? 0.95;
   const initialNoiseFloor = energyParams.initialNoiseFloor ?? 1e-3;
   const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.05;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 5e-3;
-  const minSNR = energyParams.minSNR ?? 6;
-  const snrRange = energyParams.snrRange ?? 12;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.01;
+  const minSNR = energyParams.minSNR ?? 10;
+  const snrRange = energyParams.snrRange ?? 10;
   const minEnergy = energyParams.minEnergy ?? 5e-4;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
@@ -176,9 +176,21 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
       this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
     } else {
       // If signal is louder, adapt upwards
-      // If we are currently speaking, adapt EXTREMELY slowly to avoid "chasing" speech
-      // If we are silent, adapt at the normal loud rate
-      const adaptRate = this.isSpeaking ? (this.noiseFloorAdaptRateLoud * 0.02) : this.noiseFloorAdaptRateLoud;
+      // 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));
+      
+      let multiplier = 1.0;
+      if (this.isSpeaking) {
+        multiplier = 0.01;
+      } else if (snrDb > 20) {
+        multiplier = 0.1;
+      }
+      
+      const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
       this.noiseFloor = this.noiseFloor * (1 - adaptRate) + instantRms * adaptRate;
     }
 
@@ -311,11 +323,11 @@ var VADStateMachine = class {
       enabled: config?.enabled ?? true,
       pluginName: config?.pluginName ?? "energy-vad",
       // Voice-optimized defaults
-      startThreshold: config?.startThreshold ?? 0.6,
+      startThreshold: config?.startThreshold ?? 0.8,
       // Higher threshold to avoid noise
-      stopThreshold: config?.stopThreshold ?? 0.45,
+      stopThreshold: config?.stopThreshold ?? 0.3,
       // Balanced for voice
-      hangoverMs: config?.hangoverMs ?? 400,
+      hangoverMs: config?.hangoverMs ?? 300,
       // Smooth for natural speech
       preRollMs: config?.preRollMs ?? 250,
       // Generous pre-roll
@@ -325,9 +337,9 @@ var VADStateMachine = class {
         smoothing: config?.energyVad?.smoothing ?? 0.95,
         initialNoiseFloor: config?.energyVad?.initialNoiseFloor ?? 1e-3,
         noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 0.05,
-        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 5e-3,
-        minSNR: config?.energyVad?.minSNR ?? 6,
-        snrRange: config?.energyVad?.snrRange ?? 12,
+        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.01,
+        minSNR: config?.energyVad?.minSNR ?? 10,
+        snrRange: config?.energyVad?.snrRange ?? 10,
         minEnergy: config?.energyVad?.minEnergy ?? 5e-4
       }
     };

package/dist/livekit/integration.mjs

@@ -1,12 +1,12 @@
 import {
   attachProcessingToTrack
-} from "../chunk-UQG6Z5W3.mjs";
-import "../chunk-JP6DA62Y.mjs";
-import "../chunk-2EX3FXSF.mjs";
+} from "../chunk-Y6IG7XGC.mjs";
+import "../chunk-2UPI6VWY.mjs";
+import "../chunk-XHMNP7NC.mjs";
 import "../chunk-OZ7KMC4S.mjs";
-import "../chunk-BMVZ3KKG.mjs";
+import "../chunk-FOGC2MFA.mjs";
 import "../chunk-XO6B3D4A.mjs";
-import "../chunk-2TKYGFMC.mjs";
+import "../chunk-3A2CTC4K.mjs";
 export {
   attachProcessingToTrack
 };

package/dist/pipeline/audio-pipeline.js

@@ -126,9 +126,9 @@ var createEnergyVadWorkletCode = (vadConfig) => {
   const smoothing = energyParams.smoothing ?? 0.95;
   const initialNoiseFloor = energyParams.initialNoiseFloor ?? 1e-3;
   const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.05;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 5e-3;
-  const minSNR = energyParams.minSNR ?? 6;
-  const snrRange = energyParams.snrRange ?? 12;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.01;
+  const minSNR = energyParams.minSNR ?? 10;
+  const snrRange = energyParams.snrRange ?? 10;
   const minEnergy = energyParams.minEnergy ?? 5e-4;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
@@ -174,9 +174,21 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
       this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
     } else {
       // If signal is louder, adapt upwards
-      // If we are currently speaking, adapt EXTREMELY slowly to avoid "chasing" speech
-      // If we are silent, adapt at the normal loud rate
-      const adaptRate = this.isSpeaking ? (this.noiseFloorAdaptRateLoud * 0.02) : this.noiseFloorAdaptRateLoud;
+      // 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));
+      
+      let multiplier = 1.0;
+      if (this.isSpeaking) {
+        multiplier = 0.01;
+      } else if (snrDb > 20) {
+        multiplier = 0.1;
+      }
+      
+      const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
       this.noiseFloor = this.noiseFloor * (1 - adaptRate) + instantRms * adaptRate;
     }
 
@@ -309,11 +321,11 @@ var VADStateMachine = class {
       enabled: config?.enabled ?? true,
       pluginName: config?.pluginName ?? "energy-vad",
       // Voice-optimized defaults
-      startThreshold: config?.startThreshold ?? 0.6,
+      startThreshold: config?.startThreshold ?? 0.8,
       // Higher threshold to avoid noise
-      stopThreshold: config?.stopThreshold ?? 0.45,
+      stopThreshold: config?.stopThreshold ?? 0.3,
       // Balanced for voice
-      hangoverMs: config?.hangoverMs ?? 400,
+      hangoverMs: config?.hangoverMs ?? 300,
       // Smooth for natural speech
       preRollMs: config?.preRollMs ?? 250,
       // Generous pre-roll
@@ -323,9 +335,9 @@ var VADStateMachine = class {
         smoothing: config?.energyVad?.smoothing ?? 0.95,
         initialNoiseFloor: config?.energyVad?.initialNoiseFloor ?? 1e-3,
         noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 0.05,
-        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 5e-3,
-        minSNR: config?.energyVad?.minSNR ?? 6,
-        snrRange: config?.energyVad?.snrRange ?? 12,
+        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.01,
+        minSNR: config?.energyVad?.minSNR ?? 10,
+        snrRange: config?.energyVad?.snrRange ?? 10,
         minEnergy: config?.energyVad?.minEnergy ?? 5e-4
       }
     };

package/dist/pipeline/audio-pipeline.mjs

@@ -1,11 +1,11 @@
 import {
   createAudioPipeline
-} from "../chunk-JP6DA62Y.mjs";
-import "../chunk-2EX3FXSF.mjs";
+} from "../chunk-2UPI6VWY.mjs";
+import "../chunk-XHMNP7NC.mjs";
 import "../chunk-OZ7KMC4S.mjs";
-import "../chunk-BMVZ3KKG.mjs";
+import "../chunk-FOGC2MFA.mjs";
 import "../chunk-XO6B3D4A.mjs";
-import "../chunk-2TKYGFMC.mjs";
+import "../chunk-3A2CTC4K.mjs";
 export {
   createAudioPipeline
 };

package/dist/types.d.mts

@@ -43,7 +43,7 @@ interface AudioProcessingConfig {
          * When VAD probability rises above this, audio is unmuted.
          * Lower = more sensitive (catches quiet speech, may include noise)
          * Higher = less sensitive (only confident speech, may clip quiet parts)
-         * Default: 0.6 (optimized for voice-only)
+         * Default: 0.8 (aggressive noise rejection)
          */
         startThreshold?: number;
         /**
@@ -51,7 +51,7 @@ interface AudioProcessingConfig {
          * When VAD probability drops below this (after hangover), audio is muted.
          * Lower = keeps audio on longer (less aggressive gating)
          * Higher = mutes faster (more aggressive noise suppression)
-         * Default: 0.45 (balanced voice detection)
+         * Default: 0.3 (wide hysteresis for stability)
          */
         stopThreshold?: number;
         /**
@@ -59,7 +59,7 @@ interface AudioProcessingConfig {
          * Prevents rapid on/off toggling during pauses.
          * Lower = more aggressive gating, may clip between words
          * Higher = smoother but may let trailing noise through
-         * Default: 400ms (optimized for natural speech)
+         * Default: 300ms
          */
         hangoverMs?: number;
         /**
@@ -102,17 +102,17 @@ interface AudioProcessingConfig {
             noiseFloorAdaptRateQuiet?: number;
             /**
              * Rate at which noise floor adapts to loud signals (0-1).
-             * Default: 0.005 (slower adaptation for speech)
+             * Default: 0.01 (faster tracking of rising noise)
              */
             noiseFloorAdaptRateLoud?: number;
             /**
              * Minimum SNR (Signal-to-Noise Ratio) in dB for speech detection.
-             * Default: 6.0 (voice is ~2x louder than noise floor)
+             * Default: 10.0 (more aggressive noise rejection)
              */
             minSNR?: number;
             /**
              * SNR range in dB for probability scaling.
-             * Default: 12.0 (probability scales from minSNR to minSNR+snrRange)
+             * Default: 10.0 (probability scales from minSNR to minSNR+snrRange)
              */
             snrRange?: number;
             /**

package/dist/types.d.ts

@@ -43,7 +43,7 @@ interface AudioProcessingConfig {
          * When VAD probability rises above this, audio is unmuted.
          * Lower = more sensitive (catches quiet speech, may include noise)
          * Higher = less sensitive (only confident speech, may clip quiet parts)
-         * Default: 0.6 (optimized for voice-only)
+         * Default: 0.8 (aggressive noise rejection)
          */
         startThreshold?: number;
         /**
@@ -51,7 +51,7 @@ interface AudioProcessingConfig {
          * When VAD probability drops below this (after hangover), audio is muted.
          * Lower = keeps audio on longer (less aggressive gating)
          * Higher = mutes faster (more aggressive noise suppression)
-         * Default: 0.45 (balanced voice detection)
+         * Default: 0.3 (wide hysteresis for stability)
          */
         stopThreshold?: number;
         /**
@@ -59,7 +59,7 @@ interface AudioProcessingConfig {
          * Prevents rapid on/off toggling during pauses.
          * Lower = more aggressive gating, may clip between words
          * Higher = smoother but may let trailing noise through
-         * Default: 400ms (optimized for natural speech)
+         * Default: 300ms
          */
         hangoverMs?: number;
         /**
@@ -102,17 +102,17 @@ interface AudioProcessingConfig {
             noiseFloorAdaptRateQuiet?: number;
             /**
              * Rate at which noise floor adapts to loud signals (0-1).
-             * Default: 0.005 (slower adaptation for speech)
+             * Default: 0.01 (faster tracking of rising noise)
              */
             noiseFloorAdaptRateLoud?: number;
             /**
              * Minimum SNR (Signal-to-Noise Ratio) in dB for speech detection.
-             * Default: 6.0 (voice is ~2x louder than noise floor)
+             * Default: 10.0 (more aggressive noise rejection)
              */
             minSNR?: number;
             /**
              * SNR range in dB for probability scaling.
-             * Default: 12.0 (probability scales from minSNR to minSNR+snrRange)
+             * Default: 10.0 (probability scales from minSNR to minSNR+snrRange)
              */
             snrRange?: number;
             /**

package/dist/vad/vad-node.js

@@ -28,9 +28,9 @@ var createEnergyVadWorkletCode = (vadConfig) => {
   const smoothing = energyParams.smoothing ?? 0.95;
   const initialNoiseFloor = energyParams.initialNoiseFloor ?? 1e-3;
   const noiseFloorAdaptRateQuiet = energyParams.noiseFloorAdaptRateQuiet ?? 0.05;
-  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 5e-3;
-  const minSNR = energyParams.minSNR ?? 6;
-  const snrRange = energyParams.snrRange ?? 12;
+  const noiseFloorAdaptRateLoud = energyParams.noiseFloorAdaptRateLoud ?? 0.01;
+  const minSNR = energyParams.minSNR ?? 10;
+  const snrRange = energyParams.snrRange ?? 10;
   const minEnergy = energyParams.minEnergy ?? 5e-4;
   return `
 class EnergyVadProcessor extends AudioWorkletProcessor {
@@ -76,9 +76,21 @@ class EnergyVadProcessor extends AudioWorkletProcessor {
       this.noiseFloor = this.noiseFloor * (1 - this.noiseFloorAdaptRateQuiet) + instantRms * this.noiseFloorAdaptRateQuiet;
     } else {
       // If signal is louder, adapt upwards
-      // If we are currently speaking, adapt EXTREMELY slowly to avoid "chasing" speech
-      // If we are silent, adapt at the normal loud rate
-      const adaptRate = this.isSpeaking ? (this.noiseFloorAdaptRateLoud * 0.02) : this.noiseFloorAdaptRateLoud;
+      // 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));
+      
+      let multiplier = 1.0;
+      if (this.isSpeaking) {
+        multiplier = 0.01;
+      } else if (snrDb > 20) {
+        multiplier = 0.1;
+      }
+      
+      const adaptRate = this.noiseFloorAdaptRateLoud * multiplier;
       this.noiseFloor = this.noiseFloor * (1 - adaptRate) + instantRms * adaptRate;
     }
 

package/dist/vad/vad-node.mjs

@@ -1,6 +1,6 @@
 import {
   EnergyVADPlugin
-} from "../chunk-2TKYGFMC.mjs";
+} from "../chunk-3A2CTC4K.mjs";
 export {
   EnergyVADPlugin
 };

package/dist/vad/vad-state.js

@@ -36,11 +36,11 @@ var VADStateMachine = class {
       enabled: config?.enabled ?? true,
       pluginName: config?.pluginName ?? "energy-vad",
       // Voice-optimized defaults
-      startThreshold: config?.startThreshold ?? 0.6,
+      startThreshold: config?.startThreshold ?? 0.8,
       // Higher threshold to avoid noise
-      stopThreshold: config?.stopThreshold ?? 0.45,
+      stopThreshold: config?.stopThreshold ?? 0.3,
       // Balanced for voice
-      hangoverMs: config?.hangoverMs ?? 400,
+      hangoverMs: config?.hangoverMs ?? 300,
       // Smooth for natural speech
       preRollMs: config?.preRollMs ?? 250,
       // Generous pre-roll
@@ -50,9 +50,9 @@ var VADStateMachine = class {
         smoothing: config?.energyVad?.smoothing ?? 0.95,
         initialNoiseFloor: config?.energyVad?.initialNoiseFloor ?? 1e-3,
         noiseFloorAdaptRateQuiet: config?.energyVad?.noiseFloorAdaptRateQuiet ?? 0.05,
-        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 5e-3,
-        minSNR: config?.energyVad?.minSNR ?? 6,
-        snrRange: config?.energyVad?.snrRange ?? 12,
+        noiseFloorAdaptRateLoud: config?.energyVad?.noiseFloorAdaptRateLoud ?? 0.01,
+        minSNR: config?.energyVad?.minSNR ?? 10,
+        snrRange: config?.energyVad?.snrRange ?? 10,
         minEnergy: config?.energyVad?.minEnergy ?? 5e-4
       }
     };

package/dist/vad/vad-state.mjs

@@ -1,6 +1,6 @@
 import {
   VADStateMachine
-} from "../chunk-2EX3FXSF.mjs";
+} from "../chunk-XHMNP7NC.mjs";
 export {
   VADStateMachine
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tensamin/audio",
-  "version": "0.1.7",
+  "version": "0.1.8",
   "main": "dist/index.js",
   "module": "dist/index.mjs",
   "types": "dist/index.d.ts",