openvibers/openvibe.live Commits

Transcribe: hard concurrency gate so decodes cannot pile up

OP OpenVibers committed 18 Aug 2026 · cad90600ae · 1 file, +34 −1

Observed on the box: THREE large-v3-turbo whisper processes running at once
(896MB + 636MB + 636MB RSS) with load average 20.9 on 4 cores. That starved the
live encoders and the clip cutter, and is why a batch of clip re-cuts crawled.

ai-analysis serialises VOD work through _txChain, but that only covers callers
routed through it — the timeline job, clip transcripts and the on-finalize
webhook all reach transcribeWavDetailed by other paths, so the effective
concurrency was unbounded.

Put the limit where every caller must pass: transcribeWavDetailed now acquires a
slot (WHISPER_MAX_CONCURRENT, default 1) and holds it for the whole decode,
releasing in a finally so a throw cannot leak it. Waiters are handed the slot
directly rather than re-polling.

The full-VOD coverage change made this urgent — walking an entire recording
instead of sampling three 60s windows is far more decode work per VOD, so
anything that let two of those overlap was going to saturate the box.

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

Changes

ddiff --git a/server/ai/transcribe.js b/server/ai/transcribe.js
iindex c914e848..37aeafe3 100644
---- a/server/ai/transcribe.js
++++ b/server/ai/transcribe.js
@@@ -93,6 +93,27 @@ function killActive() {
      _active.clear();
      return n;
  }
++// ── Hard concurrency gate ─────────────────────────────────────────────────
++// ai-analysis serialises VOD work through _txChain, but that only covers callers that
++// go through it — the timeline job, clip transcripts and the on-finalize webhook all
++// reach transcribeWavDetailed by other routes. Three large-v3-turbo processes were
++// observed running at once (896MB + 636MB + 636MB RSS, load average 20.9 on 4 cores),
++// which starved the live encoders and the clip cutter alike.
++//
++// Put the limit HERE, where every caller must pass, so no future path can bypass it.
++const MAX_CONCURRENT = Math.max(1, parseInt(process.env.WHISPER_MAX_CONCURRENT, 10) || 1);
++let _running = 0;
++const _waiters = [];
++function _acquire() {
++    if (_running < MAX_CONCURRENT) { _running++; return Promise.resolve(); }
++    return new Promise(resolve => _waiters.push(resolve));
++}
++function _release() {
++    const next = _waiters.shift();
++    if (next) next();              // hand the slot straight over
++    else _running = Math.max(0, _running - 1);
++}
++
  // While a stream is live we lower the whisper thread count so VOD transcription can
  // still make progress without starving the live encoders.
  let _lowPower = false;
@@@ -160,7 +181,9 @@ function _joinSegments(segments) {
   * @returns {Promise<{text:string, segments:Array<{start:number,end:number,text:string}>}>}
   *          start/end in seconds (shifted by opts.offsetSec).
   */
--function transcribeWavDetailed(wavPath, { timeoutMs = 180000, offsetSec = 0, live = false } = {}) {
++// Inner implementation. Wrapped by transcribeWavDetailed, which holds the concurrency
++// slot for the whole run so only MAX_CONCURRENT decoders exist at any moment.
++function _transcribeWavInner(wavPath, { timeoutMs = 180000, offsetSec = 0, live = false } = {}) {
      return new Promise((resolve) => {
          const bin = whisperBin();
          // ok=false only for genuine FAILURES (missing binary, spawn/exec error, timeout,
@@@ -250,4 +273,14 @@ async function transcribeMedia(mediaPath, opts = {}) {
      return r.text;
  }
  
++/**
++ * Detailed transcription of a 16kHz mono WAV, serialised against every other caller.
++ * The slot is held for the whole decode and always released, including on throw.
++ */
++async function transcribeWavDetailed(wavPath, opts = {}) {
++    await _acquire();
++    try { return await _transcribeWavInner(wavPath, opts); }
++    finally { _release(); }
++}
++
  module.exports = { available, transcribeWav, transcribeWavDetailed, transcribeMedia, transcribeMediaDetailed, killActive, setLowPower };