AI-Hear: log the captured audio level

OP OpenVibers committed 18 Aug 2026 · 2aa4eb6eaf · 1 file, +12 −0

[AI-Hear] has been logging audio=captured with transcript="" every cycle for
hours. The capture check only tests file SIZE, which proves nothing — WAV is
uncompressed, so 12s of pure silence is the same 384KB as 12s of speech.

Measure mean/max dB on the captured chunk so an empty transcript can be
attributed: silent capture is an audio-path bug, loud capture with no text is a
model problem. Diagnostic only.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Xw6Sf5zzSALw8m93XNu5JE

Changes

ddiff --git a/server/ai/stream-audio.js b/server/ai/stream-audio.js
iindex 728dcb38..e8e9899c 100644
---- a/server/ai/stream-audio.js
++++ b/server/ai/stream-audio.js
@@@ -142,6 +142,18 @@ async function captureWebrtc(stream, seconds) {
      cleanup();
      const size = (ok && fs.existsSync(out)) ? fs.statSync(out).size : 0;
      if (size > 8000) { // >~0.25s of 16k mono; smaller = effectively empty
++        // Size proves nothing: WAV is uncompressed, so 12s of pure SILENCE is the same
++        // 384KB as 12s of speech. Measure the actual level — an empty transcript on a
++        // silent capture is an audio-path bug, on a loud capture it is a model problem,
++        // and until now we could not tell the two apart.
++        try {
++            const { execFileSync } = require('child_process');
++            const err = execFileSync('ffmpeg', ['-hide_banner', '-i', out, '-af', 'volumedetect', '-f', 'null', '-'],
++                { encoding: 'utf8', stdio: ['ignore', 'ignore', 'pipe'], timeout: 15000 });
++            const mean = /mean_volume:\s*(-?[\d.]+) dB/.exec(err);
++            const max = /max_volume:\s*(-?[\d.]+) dB/.exec(err);
++            console.log(`[AI-Hear] stream ${stream.id}: capture level mean=${mean ? mean[1] : '?'}dB max=${max ? max[1] : '?'}dB (${size} bytes)`);
++        } catch (e) { console.log(`[AI-Hear] stream ${stream.id}: volumedetect failed: ${e.message}`); }
          return out;
      }
      console.warn(`[AI-Hear] stream ${stream.id}: capture produced ${size} bytes (ffmpeg ok=${ok}) — treating as empty`);

diff --git a/server/ai/stream-audio.js b/server/ai/stream-audio.js

index 728dcb38..e8e9899c 100644

--- a/server/ai/stream-audio.js

+++ b/server/ai/stream-audio.js

@@ -142,6 +142,18 @@ async function captureWebrtc(stream, seconds) {

cleanup();

const size = (ok && fs.existsSync(out)) ? fs.statSync(out).size : 0;

if (size > 8000) { // >~0.25s of 16k mono; smaller = effectively empty

+ // Size proves nothing: WAV is uncompressed, so 12s of pure SILENCE is the same

+ // 384KB as 12s of speech. Measure the actual level — an empty transcript on a

+ // silent capture is an audio-path bug, on a loud capture it is a model problem,

+ // and until now we could not tell the two apart.

+ try {

+ const { execFileSync } = require('child_process');

+ const err = execFileSync('ffmpeg', ['-hide_banner', '-i', out, '-af', 'volumedetect', '-f', 'null', '-'],

+ { encoding: 'utf8', stdio: ['ignore', 'ignore', 'pipe'], timeout: 15000 });

+ const mean = /mean_volume:\s*(-?[\d.]+) dB/.exec(err);

+ const max = /max_volume:\s*(-?[\d.]+) dB/.exec(err);

+ console.log(`[AI-Hear] stream ${stream.id}: capture level mean=${mean ? mean[1] : '?'}dB max=${max ? max[1] : '?'}dB (${size} bytes)`);

+ } catch (e) { console.log(`[AI-Hear] stream ${stream.id}: volumedetect failed: ${e.message}`); }

return out;

}

console.warn(`[AI-Hear] stream ${stream.id}: capture produced ${size} bytes (ffmpeg ok=${ok}) — treating as empty`);