sangdth/phowhisper-large-mlx-16bit

PhoWhisper-large (MLX)

vinai/PhoWhisper-large converted to MLX for mlx-whisper on Apple Silicon. The weights are unchanged, stored as float16. All credit for the model goes to VinAI; see PhoWhisper's homepage.

Use

import mlx_whisper

result = mlx_whisper.transcribe("speech.wav", path_or_hf_repo="sangdth/phowhisper-large-mlx-16bit", language="vi")
print(result["text"])

How it was made

convert.py from ml-explore/mlx-examples (commit 796f5b5), run on vinai/PhoWhisper-large commit b9136a4:

python convert.py --torch-name-or-path vinai/PhoWhisper-large --mlx-path phowhisper-large-mlx

The output file model.safetensors is renamed to weights.safetensors, the name mlx-whisper loads.

Citation

@inproceedings{PhoWhisper,
  title     = {{PhoWhisper: Automatic Speech Recognition for Vietnamese}},
  author    = {Thanh-Thien Le and Linh The Nguyen and Dat Quoc Nguyen},
  booktitle = {Proceedings of the ICLR 2024 Tiny Papers track},
  year      = {2024}
}

License

BSD 3-Clause, Copyright (c) 2024, VinAI. The full text is in LICENSE.