per2021/whisper-bsc-large-v3-cat-mlx

whisper-bsc-large-v3-cat, converted to MLX

BSC-LT/whisper-bsc-large-v3-cat converted to the MLX format so it runs on the GPU of Apple silicon through mlx-whisper. The weights are the same model, in float16; nothing was retrained or pruned.

Use

pip install mlx-whisper

import mlx_whisper

result = mlx_whisper.transcribe(
    "audio.wav",
    path_or_hf_repo="per2021/whisper-bsc-large-v3-cat-mlx",
    language="ca",
)
print(result["text"])

Why it exists

There was no MLX build of this model, so anyone on a Mac had to download the PyTorch weights and convert them by hand.

For accuracy, see the evaluation in the original model card, measured on Catalan test sets.

Two practical notes:

• Its transcripts come out normalised: no punctuation, no casing.
• Being a full large-v3, it is slower than whisper-large-v3-turbo. Whisper always processes a 30-second window, so the cost per utterance does not depend on how long you speak.

Conversion

Apple's own script, with one caveat: it writes model.safetensors while mlx-whisper looks for weights.safetensors.

curl -sL -o convert.py \
  https://raw.githubusercontent.com/ml-explore/mlx-examples/main/whisper/convert.py
python convert.py \
  --torch-name-or-path BSC-LT/whisper-bsc-large-v3-cat \
  --mlx-path whisper-ca-bsc-mlx \
  --dtype float16
mv whisper-ca-bsc-mlx/model.safetensors whisper-ca-bsc-mlx/weights.safetensors

Credits and licence

• Fine-tune: Barcelona Supercomputing Center, trained on 3CatParla, Parlament Parla, Corts Valencianes and Common Voice.
• Base model: openai/whisper-large-v3.
• Licence: Apache 2.0, inherited from the original. This derivative only changes the weight format.

Related

• per2021/whisper-large-v3-ca-punctuated-mlx: a sibling fine-tune whose transcripts come out with question marks, commas and casing. Same file size, and the same speed in our testing.