@absolutejs/voice 0.0.22-beta.611 → 0.0.22-beta.613

package/dist/core/bargeInDetector.d.ts

@@ -15,6 +15,14 @@ export type VoiceBargeInVerdict = {
     shouldCancel: boolean;
     /** Diagnostic label, surfaced on the barge_in / barge_in_suppressed trace. */
     reason?: string;
+    /**
+     * The acoustic measurements the decision used, surfaced on the trace for
+     * tuning the thresholds against real audio. Omitted when no audio was judged.
+     */
+    metrics?: {
+        voicedMs: number;
+        rms: number;
+    };
 };
 export type VoiceBargeInDetector = {
     evaluate: (input: VoiceBargeInInput) => Promise<VoiceBargeInVerdict> | VoiceBargeInVerdict;

package/dist/index.js

@@ -5287,12 +5287,17 @@ var createVoiceSession = (options) => {
             partialText: triggeringText,
             wordCount,
             ...getTurnAudioForDetector()
-          })) : { reason: undefined, shouldCancel: !isBackchannelByText };
+          })) : {
+            metrics: undefined,
+            reason: undefined,
+            shouldCancel: !isBackchannelByText
+          };
           const reason = verdict.reason ?? (verdict.shouldCancel ? "stt_partial" : "backchannel");
           if (verdict.shouldCancel) {
             backchannelSuppressedAt = null;
             appendTurnLatencyStage({
               metadata: {
+                ...verdict.metrics,
                 partial: triggeringText.slice(0, 200),
                 source: reason,
                 wordCount
@@ -5305,6 +5310,7 @@ var createVoiceSession = (options) => {
             backchannelSuppressedAt = Date.now();
             appendTurnLatencyStage({
               metadata: {
+                ...verdict.metrics,
                 partial: triggeringText.slice(0, 200),
                 reason,
                 wordCount
@@ -42001,26 +42007,29 @@ var createRegexSemanticTurnDetector = (options) => {
   };
 };
 // src/core/bargeInDetector.ts
+var VOICED_FLOOR = 0.02;
 var measureTurnAudio = (chunks, format) => {
   const channels = format.channels ?? 1;
   const sampleRate = format.sampleRateHz ?? 16000;
   let sumSquares = 0;
-  let sampleCount = 0;
+  let voicedSamples = 0;
   for (const chunk of chunks) {
     const usableBytes = chunk.byteLength - chunk.byteLength % 2;
     const view = new DataView(chunk.buffer, chunk.byteOffset, usableBytes);
     for (let offset = 0;offset < usableBytes; offset += 2) {
       const sample = view.getInt16(offset, true) / 32768;
-      sumSquares += sample * sample;
-      sampleCount += 1;
+      if (Math.abs(sample) >= VOICED_FLOOR) {
+        sumSquares += sample * sample;
+        voicedSamples += 1;
+      }
     }
   }
-  if (sampleCount === 0) {
+  if (voicedSamples === 0) {
     return { durationMs: 0, rms: 0 };
   }
   return {
-    durationMs: sampleCount / channels / sampleRate * 1000,
-    rms: Math.sqrt(sumSquares / sampleCount)
+    durationMs: voicedSamples / channels / sampleRate * 1000,
+    rms: Math.sqrt(sumSquares / voicedSamples)
   };
 };
 var createAcousticBargeInDetector = (options = {}) => {
@@ -42034,19 +42043,20 @@ var createAcousticBargeInDetector = (options = {}) => {
         return input.isBackchannelByText ? { reason: "text_backchannel", shouldCancel: false } : { reason: "text_only", shouldCancel: true };
       }
       const { durationMs, rms } = measureTurnAudio(turnAudio, turnAudioFormat);
+      const metrics = { rms, voicedMs: Math.round(durationMs) };
       if (durationMs >= sustainedMs) {
-        return { reason: "acoustic_sustained", shouldCancel: true };
+        return { metrics, reason: "acoustic_sustained", shouldCancel: true };
       }
       if (input.isBackchannelByText) {
-        return { reason: "acoustic_backchannel", shouldCancel: false };
+        return { metrics, reason: "acoustic_backchannel", shouldCancel: false };
       }
       if (rms >= emphaticRms) {
-        return { reason: "acoustic_emphatic", shouldCancel: true };
+        return { metrics, reason: "acoustic_emphatic", shouldCancel: true };
       }
       if (rms <= noiseFloorRms) {
-        return { reason: "acoustic_noise_floor", shouldCancel: false };
+        return { metrics, reason: "acoustic_noise_floor", shouldCancel: false };
       }
-      return { reason: "acoustic_ambiguous", shouldCancel: true };
+      return { metrics, reason: "acoustic_ambiguous", shouldCancel: true };
     }
   };
 };

package/dist/testing/index.js

@@ -7514,12 +7514,17 @@ var createVoiceSession = (options) => {
             partialText: triggeringText,
             wordCount,
             ...getTurnAudioForDetector()
-          })) : { reason: undefined, shouldCancel: !isBackchannelByText };
+          })) : {
+            metrics: undefined,
+            reason: undefined,
+            shouldCancel: !isBackchannelByText
+          };
           const reason = verdict.reason ?? (verdict.shouldCancel ? "stt_partial" : "backchannel");
           if (verdict.shouldCancel) {
             backchannelSuppressedAt = null;
             appendTurnLatencyStage({
               metadata: {
+                ...verdict.metrics,
                 partial: triggeringText.slice(0, 200),
                 source: reason,
                 wordCount
@@ -7532,6 +7537,7 @@ var createVoiceSession = (options) => {
             backchannelSuppressedAt = Date.now();
             appendTurnLatencyStage({
               metadata: {
+                ...verdict.metrics,
                 partial: triggeringText.slice(0, 200),
                 reason,
                 wordCount

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@absolutejs/voice",
-  "version": "0.0.22-beta.611",
+  "version": "0.0.22-beta.613",
   "description": "Voice primitives and Elysia plugin for AbsoluteJS",
   "repository": {
     "type": "git",