interesting since you last looked at it assembly ai currently supports both speaker diarization

and a higher level speaker identification feature it can turn generic speaker a b labels into names

or roles when given appropriate context there is an important caveat for bare eye though assembly

ai's ordinary diarization does not intrinsically create a persistent global speaker identity across

unrelated files their own guidance for cross-file recognition is essentially the same architecture

you already have diarize the recording extract speaker embeddings and compare those against a

known speaker embedding store so assembly ai can probably replace whisper plus pi note for

transcription and within recording diarization it does not automatically make qdran based

persistent voice recognition unnecessary what i would change in beer i would move away from

treating each tiny part of its embedding as an independent recognition attempt suppose today's

recording produces open quote speaker zero zero speaker zero zero speaker zero speaker zero one

speaker zero zero close quote don't try to identify every one of those five utterances

separately instead

diarize the whole recording