The whisper-pionote mismatch you noticed is real.

Pionote might decide, speaker A, while whisper might return.

I think we should probably...

That whisper segment crosses the speaker boundary,

so you cannot safely say this whisper segment belongs to whichever pionote speaker overlaps it most

and assume it is correct.

If beer eye is doing something along those lines,

that could explain quite a lot of the odd speaker attribution we saw today.

The better representation is word-level timing.

You align the ASR words against the diarization timeline

and assign each word or groups of contiguous words to the speaker occupying that interval,

then rebuild utterances afterwards.

DeepGram actually makes this particularly straightforward

because its current diarization output assigns a speaker to each word.

Its current batch diarizer is V2.

Assembly AI similarly returns diarized utterances

and supports expected speaker counts ranges,

which can improve separation when you know approximately how many people are present.

Assembly AI absolutely does speaker diarization,

and it has become rather more...