Deepgram Flux STT

Deepgram Flux is a turn-aware speech-to-text model that decides when the caller has finished speaking. Learn when to use it and how to tune its end-of-turn thresholds.

2 min readUpdated

Deepgram Flux is a speech-to-text model that also decides when the caller has finished their turn. It runs on the default Edesy runtime as a separate provider from Deepgram Nova, so existing Deepgram agents are not affected. Use Flux when you want the end of a turn to be decided by the speech model instead of by silence alone.

At a glance#

Item Value
Provider ID deepgram_flux (alias flux)
Runtime Default runtime
Models flux-general-en for English agents, flux-general-multi for other languages
Language hints Hindi, Spanish, French, German, Portuguese; other languages are auto-detected
End-of-turn threshold Default 0.7, allowed range 0.5 to 0.9
Eager end-of-turn Off by default
Your own key (BYOK) Uses your Deepgram key when you add one

When to use Flux#

  • Callers pause mid-sentence ("my order number is... 4 5 2...") and a silence timer cuts them off.
  • You want fewer interruptions without making the agent wait longer after every turn.
  • Your agent is English or one of the hinted languages above.

Keep Deepgram Nova if your current agent works well. Switching is opt-in per agent.

Set it up#

  1. Open the agent and go to Voice → Speech-to-text.
  2. Choose Deepgram Flux. The model is picked from the agent language: English agents use flux-general-en, others use flux-general-multi.
  3. Save and place a test call. Watch how the agent handles your pauses.
  4. If the agent still cuts in too early, raise eot_threshold. If it waits too long, lower it.
{
  "sttProvider": "deepgram_flux",
  "sttConfig": {
    "eot_threshold": 0.7
  }
}

Settings#

Setting Default What it does
eot_threshold 0.7 How confident the model must be that the turn ended. Higher waits longer. Values outside 0.5 to 0.9 are clamped
eager_eot_threshold Off Starts preparing a reply early when the model thinks the turn is probably over. Must be below eot_threshold. It increases LLM calls, so it is off by default

FAQs#

Is Flux the same as Deepgram Nova? No. It uses a different endpoint and message format, which is why it is a separate provider ID.

Does Flux work for Hindi? Yes, through flux-general-multi with a Hindi hint. Test it on real calls and compare with Nova or Sarvam.

Should I turn on eager end-of-turn? Only after testing. It can make replies feel quicker, but it uses more LLM calls.

Does it change pricing? Deepgram STT is included in the per-minute price. See provider costs.

Was this page helpful?