can do this. Open quote. Speaker A, Behnam Speaker B, Damun Speaker C, David. Close quote.

But you supply Assembly AI with a possible people, name, slash roles, and optionally

descriptive context about them. It uses the content and context of the conversation to

infer who is who. Their documentation explicitly describes identification as being inferred

from conversational content. For today's meeting, for example, you might submit Behnam,

TLC, son of Reza, Reza, Behnam's father, Damun, involved with TLC, David, silvery

origami tomato supplier. And Assembly AI can use things like people addressing each

other by name, what each person says, their roles, et cetera, to turn A, B, C, D into

actual names. That's potentially extremely useful for Brian, particularly because the

price is only another two cents out. But Assembly AI does not give you its speaker embeddings.

This is the important bit. Internally, yes, Assembly AI absolutely creates speaker embeddings

as part of diarization. Their own technical explanation says that they cut the speech

into chunks, build speaker embeddings, and then cluster similar embeddings to determine

speakers.

But those internal speaker embeddings are not returned.