[notjulianjaynes]

Jumping on this as a fellow idiot to ask for suggestions on having a local LLM generate a summary from a transcript with multiple speakers. How important is it that the transcript is well formatted (diarization, etc) first. My attempts have failed miserably thus far.

Edit: using a P40, whisper as ASR

[petronic]

My experience with more traditional (non-Whisper-based) diarization & transcription is that it's heavily dependent on how well the audio is isolated. In a perfect scenario (one speaker per audio channel, well-placed mics) you'll potentially see some value from it. If you potentially have scenarios where the speaker's audio is being mixed with other sounds or music, they'll often be flagged as an additional speaker (so, speaker 1 might also be speaker 7 and 9) - which can make for a less useful summary.

[notjulianjaynes]

You're saying that diarization quality is dependent on speaker isolation? I should have clarified, to my knowledge whisper does not perform that step, and whether I need to do diarization is what I'm trying to figure out. (Probably I do.) Pyannote.audio has been suggested, but I ran into some weird dependency thing I didn't feel like troubleshooting late last night, so I have not been successful in using it yet.