you as part of the API output for you to put into U-TRAP. And Assembly AI's own documentation

for persistent recognition across recordings recommends exactly what we've been discussing.

Use Assembly AI for transcription diarization. Then use a separate speaker embedding model

such as NVIDIA Titanet, and store search those embeddings yourself. So open quote, Assembly

AI produces embeddings internally, but doesn't expose those embeddings to Braille. Close

quote. That means, yes, Braille remains responsible for true cross-recording voice

recognition. Although this actually gives you a very nice architecture, I think I'd

now make Assembly AI authoritative for audio, words, timestamps, speaker, A, B, C, D, likely

actual names, for only 25,000. Then BRI independently uses the audio that Assembly AI has grouped

into speakers to perform persistent voice recognition. And this is where I'd probably

reconsider Firehawk as the embedding model. Assembly AI themselves publish a reference

architecture using NVIDIA Titanet specifically for this purpose. Diarize with Assembly AI,

generate speaker embeddings with Titanet.

Put those in a vector database and perform cross-file speaker recognition.