Which whisper model should I pick for transcription -- is `medium` a sensible default?

On my self-hosted instance I changed the whisper model to medium and set max parallel to 4, assuming a mid-size model was a good balance between the tiny default and the large ones. Transcription is slow and I am not sure the extra time is buying me anything.

Resolution: medium is close to the worst available choice: use tiny.en for English content, or large-v3-turbo when the audio is in another language. large-v3-turbo is both faster than medium and considerably more accurate, so there is no reason to sit in the middle of the size range -- you pay medium's runtime for worse output. For English audiobooks tiny.en is the practical default and is what the pipeline is tuned around; step up to large-v3-turbo only for non-English audio or when the alignment grade comes back poor. Note that transcription model choice is independent of audio transcoding, so it is not the thing to change when chasing a playback or codec problem.