sign cutoff for your particular embedding model. I would definitely test Assembly AI now. The cost

has fallen enough that this is no longer the extravagant solution you might remember. Assembly

AI currently quotes roughly 0.20 once per hour for universal 3.5 pro async transcription with

diarization at about plus hot hour. So even 50 hours a week would only be around 0.23 than 50

but at $1.50 a week or roughly $50 per month. That is cheap enough that for beer eye I wouldn't

choose Whisper merely because it is free if Assembly AI produces materially better evidence.

The engineering time and bad attribution are worth more than that. Deepgram is also inexpensive.

Its current Nova 3 pricing is around the same order of magnitude and speaker diarization is

supported directly. Between Deepgram and Assembly AI, I'd initially favor Assembly AI.

Not because Deepgram is bad, quite the opposite, but because your workload is mostly recorded

conversations requiring forensic quality transcripts, not ultra low latency voice agent

speech recognition.

Assembly AI is a great tool for you.

Assembly AI's current feature set seems unusual.