deepgram flux

[Rej1sky]

Hey I noticed that deepgram flux doesn't work yet while Nova 3 does. I would prefer to use flux. I think it doesn't work because of some changes they made. Would be really happy it you could ad support.
Greetings
Jan

[DevEmperor]

Hi Jan, thanks for flagging this — I dug into it, and the short version is that Flux is not a drop-in replacement for Nova-3: it is streaming-only and lives on a different API version, wss://api.deepgram.com/v2/listen, with its own message format (TurnInfo with StartOfTurn / Update / EndOfTurn events instead of Results + is_final). There is no pre-recorded endpoint for it at all, so it can never work as the transcription model — only as the real-time one, and our Deepgram streaming session currently talks /v1 only. That is why Nova-3 works and Flux does not.

It is a good fit for dictation though, so I want to support it. I opened #291 with the research and a concrete plan (a separate Flux session on /v2/listen, picking it by model id, plus filtering streaming-only models out of the transcription model list where they can only fail). Let's continue there — there is also a question for you in the issue about where exactly you entered the model, which will help me start at the right end.

[DevEmperor]

Done — Flux is supported now (#291, shipped in ea9fb0a1, in the next release). Pick it in the provider settings under Real-time model; it only works there, since Flux has no pre-recorded endpoint. flux-general-en is English-only, flux-general-multi covers ten languages and switches between them mid-sentence. Thanks again for the nudge, Jan.