Deepgram flux-general-en STT throws 400, while nova-3 works

Agents
,
1

I’m trying to use the flux-general-en STT by passing this to AgentSessionOptions (node SDK):

{
stt: new deepgram.STT({ model: ‘flux-general-en’ }),
};

I’m getting the following error:

ERROR (27220): AgentSession is closing due to an unrecoverable error
type: “stt_error”
label: “deepgram.STT”
recoverable: false
error: {
“type”: “APIStatusError”,
“message”: “Deepgram WebSocket connection rejected with status 400”,
“stack”:
APIStatusError: Deepgram WebSocket connection rejected with status 400
at WebSocket.onUnexpectedResponse (C:\Projects\myproject\livekit-agent\node_modules.pnpm@livekit+agents@1.8.1_@live_1f2f4d1578e9959998da1e2ff02a2afe\node_modules@livekit\agents\src\utils.ts:69:9)
at Object.onceWrapper (node:events:623:26)
at WebSocket.emit (node:events:508:28)
at ClientRequest. (C:\Projects\myproject\livekit-agent\node_modules.pnpm\ws@8.21.0\node_modules\ws\lib\websocket.js:929:27)
at ClientRequest.emit (node:events:508:28)
at HTTPParser.parserOnIncomingClient (node:_http_client:780:27)
at HTTPParser.parserOnHeadersComplete (node:_http_common:125:17)
at TLSSocket.socketOnData (node:_http_client:615:22)
at TLSSocket.emit (node:events:508:28)
at addChunk (node:internal/streams/readable:563:12)
“body”: null,
“retryable”: false,
“name”: “APIStatusError”,
“statusCode”: 400,
“requestId”: null
}

To my understanding it should be supported:

It works as expected when passing nova-3.

2

Hi, for Flux models, please use STTv2, i.e.

import * as deepgram from '@livekit/agents-plugin-deepgram';

const session = new voice.AgentSession({
    stt: new deepgram.STTv2({
        model: "flux-general-en",
    }),
    // ... llm, tts, etc.
});
3

Got it, thanks.

Is there anything to gain from using flux instead of nova STT for transcription alongside a realtime model?

According to the comparison, the main added value seems to be turn-based transcription, but I’m not sure whether it adds anything in this configuration.

4

So you have a realtime (speech-to-speech) model in your agent and you are using a separate STT for accurate transcription? If that is the case, I agree with you that you wouldn’t gain anything by using Flux since, as you say, the primary reason to choose it is for its turn taking model.

5

Correct. Fair enough, wanted to make sure I wasn’t missing anything. Thanks Darryn.