whisper-large-v3-ca-mlx

MLX (Apple Silicon, fp16) conversion of projecte-aina/whisper-large-v3-ca-3catparla — a Whisper Large v3 fine-tuned on 710 hours of Catalan broadcast audio from the 3CatParla corpus.

Runs natively on Apple Silicon GPUs via mlx-whisper.

Model details

Property	Value
Base model	openai/whisper-large-v3
Fine-tune	projecte-aina/whisper-large-v3-ca-3catparla
Language	Catalan (ca)
Format	MLX fp16
Size	~2.9 GB
Architecture	Whisper Large v3 (32 layers, 1280 state, 20 heads)

Usage

pip install mlx-whisper

import mlx_whisper

result = mlx_whisper.transcribe(
    "audio.wav",
    path_or_hf_repo="caballosanexx/whisper-large-v3-ca-mlx",
    language="ca",
    fp16=True,
)
print(result["text"])

Supported formats: .mp3 .wav .m4a .ogg .flac

Requirements

• Apple Silicon Mac (M1 or later)
• macOS 13.3+
• mlx-whisper

Attribution

This is an MLX format conversion of the original model. All credit for training goes to:

• Authors: Carlos Daniel Hernández Mena, Carme Armentano Oller, Sarah Solito, Baybars Külebi
• Organization: Language Technologies Unit, Barcelona Supercomputing Center (BSC-LT)
• Funded by: Generalitat de Catalunya (Projecte AINA)
• Contact: langtech@bsc.es

If you use this model in research, please cite the original work:

@inproceedings{hernandez20243catparla,
  title={3CatParla: A New Open-Source Corpus of Broadcast TV in Catalan for Automatic Speech Recognition},
  author={Hernández Mena, Carlos Daniel and Armentano Oller, Carme and Solito, Sarah and Külebi, Baybars},
  booktitle={Proc. IberSPEECH 2024},
  pages={176--180},
  year={2024}
}

License

Apache 2.0 — inherited from projecte-aina/whisper-large-v3-ca-3catparla.