Whisper's prepare_audio raises LibsndfileError on WebM/MP4 uploads and unloads the whisper worker

[jeffreywangcf]

Environment

• OpenArc 2.0.5 (e5edf42-dirty), image ghcr.io/searchsavior/openarc-battlemage:latest
• Kernel 6.19.12-200.fc43.x86_64, Arc Pro B50 single dGPU exposed to the container
• Loaded models: whisper-large-v3 (fp16 IR, engine ovgenai, device GPU) and kokoro (fp16 IR, engine openvino, device GPU)
• Clients: OpenWebUI 0.10.2 voice chat and OpenWhispr, both pointed at http://<host>:4001/v1 (OpenArc running in Docker on <host>)

What I was doing

Adding OpenWhispr dictation on top of a working OpenWebUI voice-chat setup. OpenWhispr (STT client) got:

External: 400, message='Bad Request', url='http://<host>:4001/v1/audio/transcriptions'

Subsequent requests to the same endpoint returned {"detail":"Model 'whisper-large-v3' is not loaded"}.

Triage

Monitored the container while triggering one dictation:

docker logs -f openarc 2>&1 | grep -iE "whisper|libsndfile|unload"

The dictation produced:

2026-07-22 08:09:18,830 - ERROR - Whisper inference failed!
  File "/app/src/server/worker_registry.py", line 173, in infer_whisper
    async for item in whisper_model.transcribe(packet.gen_config)
  File "/app/src/engine/ov_genai/whisper.py", line 44, in transcribe
  File "/app/src/engine/ov_genai/whisper.py", line 34, in prepare_audio
    audio, sr = librosa.load(audio_buffer, sr=16000, mono=True)
  ...
soundfile.LibsndfileError: Error opening <_io.BytesIO object at 0x7fecc199dee0>: Format not recognised.
2026-07-22 08:09:18,834 - ERROR - [Whisper Worker: whisper-large-v3] Inference failed, triggering model unload...

Cause

prepare_audio in src/engine/ov_genai/whisper.py decodes uploads via librosa.load, which does not support MP4 or WebM, so a client uploading either format errors out and unloads the model until it is manually reloaded. This also breaks other services sharing the endpoint.

Possible Fix

• install ffmpeg in the Dockerfile
• in prepare_audio, fall back to use ffmpeg when librosa.load fails.

[jeffreywangcf]

I add a simple fallback to ffmpeg locally and it was able to resolve the issue. Happy to open a PR if useful.