Is there a way to implement a Timeout / cancellation for transcribe() method, espeicially model.decode()? (large-v3 sometimes stalls >200s on ~10s audio)

For certain short audio clips (~10 s), mlx-whisper (model: large-v3) spends an extremely long time inside model.decode() when called via whisper.transcribe(). In one case, I set a ~10 s application-level timeout, but logging shows the decode call continued for ~200 s before returning.

I’d like a way to cancel / time out decoding so my application can skip problematic audio and continue processing subsequent chunks.

Is there any recommended way to achieve this? Or would this require modifying the decoding loop itself?

Thank you

Start by tracing the transcribe() call path into model.decode(), focusing on the decoding loop mentioned in the report. Determine how a timeout or cancellation should return control so processing can continue with later audio chunks, then verify that the stalled large-v3 case exits as expected.