[WhisperAdapter] Restarting in ${delay}ms (attempt ${restart

Error message

[WhisperAdapter] Restarting in ${delay}ms (attempt ${restartAttempts}/${MAX_RESTARTS})

What it means

WhisperAdapter supervises its transcription worker with the same bounded-restart policy as Kokoro: on worker failure scheduleRestart() waits RESTART_DELAY_MS × attempt and calls ensureWorker() again, warning per attempt. At MAX_RESTARTS it logs an error and sets state 'terminated' so callers can detect the dead adapter rather than waiting in 'error' forever. A successful restart clears the counter (onSuccess).

Source

Thrown at packages/stage-ui/src/libs/inference/adapters/whisper.ts:185

      messageListener = null
      errorListener = null
      worker.terminate()
      worker = null
    }
  }

  function scheduleRestart(): void {
    if (restartAttempts >= MAX_RESTARTS) {
      console.error(`[WhisperAdapter] Max restart attempts (${MAX_RESTARTS}) reached.`)
      // NOTICE: Transition to 'terminated' so callers can detect the dead adapter
      // instead of being stuck in 'error' state indefinitely.
      state = 'terminated'
      return
    }

    restartAttempts++
    const delay = RESTART_DELAY_MS * restartAttempts
    console.warn(`[WhisperAdapter] Restarting in ${delay}ms (attempt ${restartAttempts}/${MAX_RESTARTS})`)

    setTimeout(() => {
      ensureWorker()
    }, delay)
  }

  function onSuccess(): void {
    restartAttempts = 0
  }

  function ensureWorker(): Worker {
    if (!worker) {
      worker = new Worker(workerUrl, { type: 'module' })
      messageListener = (event: MessageEvent) => {
        const data = event.data
        // Forward unified protocol messages to subscribers
        if (data.type === 'progress') {
          const evt: WhisperEvent = { type: 'progress', payload: data.payload }

View on GitHub (pinned to 677329427f)

Solutions

Capture the worker's own error in devtools before this warn to find the root cause.
Reduce concurrent model loads (the GPU coordinator / load queue exists for this) to avoid OOM.
Fix worker bundling if ensureWorker fails instantly (check the worker URL loads standalone).
After 'terminated', obtain a fresh adapter (the getKokoroAdapter-style singleton reset) instead of retrying the dead one.

Defensive patterns

Strategy: retry

Prevention

Reset restartAttempts on success (onSuccess) so transient crashes do not accumulate toward termination.
Check state === 'terminated' before invoking the adapter and recreate it when needed.
Fix worker bundling issues before shipping: a 404 worker script produces instant restart loops.

When it happens

Trigger: Worker crash while loading or running the Whisper model (WASM abort, OOM), onnxruntime-web failures under webgpu execution provider, or the worker script failing to instantiate (404/bundle error).

Common situations: Memory-constrained tabs running ASR + TTS together; bundler misconfig serving a broken worker chunk; version skew between transformers.js and the whisper ONNX artifacts.

Related errors

AI-assisted analysis of moeru-ai/airi@677329427f (2026-08-18). Data as JSON: /api/errors/7e967d1076eca91d. Report an issue: GitHub.