Add MLX Whisper Backend for Apple Silicon

[QuentinFuxa]

Hi,

As indicated in my previous pull request, this PR concerns exclusively the mlx-whisper backend:

The Whisper MLX backend is optimized specifically for Apple Silicon. It uses the mlx-whisper library, which leverages the native performance capabilities of Apple hardware for faster transcription without requiring a dedicated GPU.

Key Changes

• MLXWhisper Backend: Added MLXWhisper as a new class in the backend options.

  • Automatically translates model names (e.g., large-v2) to MLX-compatible Hugging Face paths (e.g., whisper-large-v2-mlx).
  • No change to the --model parameter in the project CLI, model "translation" is done in the background.

• Readme has been updated with information & how to install (pip install mlx-whisper)

Testing

• The implementation has been tested on Apple M1. It works faster than faster whisper.

Usage Example

python whisper_streaming.py --model "large-v3-turbo" --backend "mlx-whisper"

The MLXWhisper class translates the model name to the Hugging Face path mlx-community/whisper-large-v3-turbo and loads the model efficiently.

Let me know if you have any questions or need further clarifications!

Best regards,
Quentin

[Gldkslfmsd]

OK, thanks. I'm changing my mind. I can't review and test this code but I'd like to ask any volunteer to do so and let me know. Then I'd merge it.