mustafaaljadery/lightning-whisper-mlx ? reverse-engineered prompt

Build me a fast Whisper transcription library for Apple Silicon that feels easy to use from Python.

I want to be able to load a model, point it at an audio file, and get back the transcript as plain text. Make it optimized for speed on Macs with Apple Silicon, with support for different Whisper model sizes, batch decoding for better throughput, and optional 4bit and 8bit quantization so it can run faster and use less memory. It should work as a simple package I can install and import, with a clean little API like creating a Whisper object and calling transcribe on an audio path.

Please also include a basic example script or test that shows it working, and make sure the defaults are sensible for most people, while still letting me tune batch size and model choice if needed. If anything needs checking against the latest MLX or Whisper docs, look it up online first.