That sounds almost uncannily close to what Braille already does.

Open quote, assembly AI, titanet, Q-drant, Braille person.

Close quote, rather than your present, open quote,

whisper plus piano diarization, piano embeddings,

timing reconciliation, plus Q-drant, Braille person.

Close quote.

It removes a great deal of complexity,

and you've got another valuable signal as well.

Open quote, assembly AI says this is probably Damoon titanet,

Q-drant says this is probably Damoon extremely high confidence, Damoon.

Close quote.

Whereas, open quote, assembly AI context says Damoon voice recognition

says Reza, don't automatically assign.

Something needs checking.

Close quote.

That could make the system substantially more reliable.

One other thing I discovered which directly explains some of your current trouble.

Assembly AI says it typically needs around 30 seconds of speech

from somebody across a recording

before it can robustly establish them as a distinct speaker.

Very short, yes, okay, right.

Contributions can actually get absorbed into another speaker's cluster.

So your instinct that trying to identify every individual tiny utterance independently

is making life unnecessarily hard

looks like a little bit of a waste of time.

Close quote.