whisper-large-v3-atco2-asr-mlx

whisper-large-v3-atco2-asr-mlx

MLX (float16) build of jlvdoorn/whisper-large-v3-atco2-asr for Apple Silicon inference via mlx-whisper.

The base model is an aviation-fine-tuned Whisper-large-v3 that reports ~17% WER on ATCO2 vs ~37–55% WER for stock large-v3 on the same domain. No public MLX conversion existed at conversion time (2026-04), so this repo exists to unblock MPS evaluation.

Usage

import mlx_whisper

result = mlx_whisper.transcribe(
    "audio.wav",
    path_or_hf_repo="sfabriece/whisper-large-v3-atco2-asr-mlx",
    language="en",
)
print(result["text"])

Conversion

Converted with the mlx-examples/whisper/convert.py script:

python convert.py \
  --torch-name-or-path jlvdoorn/whisper-large-v3-atco2-asr \
  --mlx-path ./whisper-large-v3-atco2-asr-mlx \
  --dtype float16
# then rename model.safetensors → weights.safetensors (mlx-whisper loader expects this name)

License

Apache-2.0 (inherited from the base model).

Attribution

• Base model: jlvdoorn/whisper-large-v3-atco2-asr — aviation fine-tune on top of OpenAI Whisper large-v3 (MIT).
• Conversion tooling: ml-explore/mlx-examples/whisper.