From cd22b1da9f3666602d93ee582af407da4dc65dd0 Mon Sep 17 00:00:00 2001
From: ValueOn AG <p.motsch@valueon.ch>
Date: Fri, 27 Feb 2026 22:14:58 +0100
Subject: [PATCH] AudioCapture: add pre-roll buffer for speech onset, fix
 diagnostic counter

Made-with: Cursor
---
 src/bot/audioCaptureProcedure.ts | 24 ++++++++++++++++++------
 1 file changed, 18 insertions(+), 6 deletions(-)

diff --git a/src/bot/audioCaptureProcedure.ts b/src/bot/audioCaptureProcedure.ts
index 95248b9..6ebb365 100644
--- a/src/bot/audioCaptureProcedure.ts
+++ b/src/bot/audioCaptureProcedure.ts
@@ -126,23 +126,30 @@ export class AudioCaptureProcedure {
             const minRmsThreshold = 0.0015;
             const samplesPerChunk = nativeRate * 2;
             const targetRate = 16000;
+            // Pre-roll: keep last 500ms of discarded silent chunks so that
+            // speech onsets at the tail of a silent window are preserved.
+            const preRollSamples = Math.ceil(nativeRate * 0.5);
 
             processor.onaudioprocess = (e: AudioProcessingEvent) => {
               const input = e.inputBuffer.getChannelData(0);
               callbackCount++;
 
               // #region agent log
+              // Count non-zero samples on EVERY callback for accurate diagnostics
+              let nonZeroThisCallback = 0;
+              for (let i = 0; i < input.length; i++) {
+                if (input[i] !== 0) nonZeroThisCallback++;
+              }
+              totalNonZeroSamples += nonZeroThisCallback;
+
               if (callbackCount <= 3 || callbackCount % 50 === 0) {
-                let nonZero = 0;
                 let maxAbs = 0;
                 for (let i = 0; i < input.length; i++) {
-                  if (input[i] !== 0) nonZero++;
                   const abs = Math.abs(input[i]);
                   if (abs > maxAbs) maxAbs = abs;
                 }
-                totalNonZeroSamples += nonZero;
                 console.log(
-                  `[AudioCapture][DIAG] onaudioprocess #${callbackCount}: bufLen=${input.length}, nonZero=${nonZero}/${input.length}, maxAbs=${maxAbs.toFixed(8)}, track.enabled=${event.track.enabled}, track.muted=${event.track.muted}, track.readyState=${event.track.readyState}, ctx.state=${ctx.state}, totalNonZero=${totalNonZeroSamples}`
+                  `[AudioCapture][DIAG] onaudioprocess #${callbackCount}: bufLen=${input.length}, nonZero=${nonZeroThisCallback}/${input.length}, maxAbs=${maxAbs.toFixed(8)}, track.enabled=${event.track.enabled}, track.muted=${event.track.muted}, track.readyState=${event.track.readyState}, ctx.state=${ctx.state}, totalNonZero=${totalNonZeroSamples}`
                 );
               }
               // #endregion
@@ -171,8 +178,13 @@ export class AudioCaptureProcedure {
                       `[AudioCapture] silent chunk skipped: track=${trackId}, readyState=${event.track.readyState}, muted=${event.track.muted}, enabled=${event.track.enabled}, rms=${rms.toFixed(6)}, callbacks=${callbackCount}, totalNonZero=${totalNonZeroSamples}`
                     );
                   }
-                  chunkBuffer = [];
-                  samplesCollected = 0;
+                  // Pre-roll: retain the tail of the silent chunk so the next
+                  // voiced chunk includes the speech onset that may have started
+                  // in the last few hundred ms of this window.
+                  const keep = Math.min(preRollSamples, merged.length);
+                  const preRoll = merged.slice(merged.length - keep);
+                  chunkBuffer = [preRoll];
+                  samplesCollected = keep;
                   return;
                 }