I have not not been having any transcriptionfidelity issues in many hours of conversation using FasterWhisper small model, so I figured I'd try out a smaller model and see if it's equally workable.

The FasterWhisper tiny model is about 75mb on my computer, which is in the ballpark of 1015 phone images worth of file size (vs ~500mb for the small model or 3gm for the large model).

The net result was: the transcriptions do miss some of what I'm saying but not in a way that Claude seemed to miss the meaning. It seemed to bring the transcription time down to ~3.5 seconds vs the ~4.5 seconds I had been seeing using the small model. I'll probably keep using the tiny model for some time and will keep an eye out for any concerns. But for now no immediate blockers.

Github: https://github.com/ccappetta/bidirect...