m-bain/whisperX ? reverse-engineered prompt

Build me a Python tool that transcribes audio files with much better timing than normal speech to text, ideally with word level timestamps and optional speaker labels for multi speaker recordings.

I want it to work from the command line on a local machine, with GPU support if available, but also a CPU mode that still works on Mac or slower laptops. It should take an audio file or folder of files, run speech detection first to skip silence, then transcribe the speech, align the words to the audio more accurately, and export a readable transcript like txt, srt, or json. If possible, add an option to mark who is speaking when there are multiple people, and let me set the number of speakers if I already know it.

Make the defaults simple so a normal user can just point it at a file and get a result, but still allow model, language, batch size, and output options to be changed. If you need to look up current docs online while wiring it up, go ahead.