collabora
Apple Silicon (MLX/CoreML) backend
WhisperLive looks like a great building block for a local voice assistant!
A lot of people interested in local AI applications (including myself) run Apple hardware due to its great price/performance in this area. Ironically faster_whisper is more like slower_whisper on M-series processors (M1/M2/M3/M4) because CTranslate2 isn't optimized falling back to the CPU.
The original ("slow") whisper.cpp has an MLX backend now, which performs better than faster_whisper on Apple processors. It would be great for it to be available to achieve more practical performance on Apple hardware.
(that said: the tiny model does achieve real time performance on M1!)
There are some other interesting options (though I'm not sure whether they fit your concept of a "backend") as well: WhisperX (which has an MLX version upcoming) supports batched inference, which, from my understanding of the process (simultaneous processing of overlapping audio snippets), could lead to further performance gains.
