Eager end-of-turn enabled by default for Deepgram Flux STT

AI Assistants using Deepgram Flux now begin generating responses before a speaker fully finishes, while still waiting for confirmed turn completion before playback. This reduces perceived latency without changing agent behavior or introducing interruptions.

What’s new

• Eager end-of-turn processing: Starts tentative LLM responses before final turn completion.
• Default behavior: Active for AI Assistants using Deepgram Flux STT.
• Configurable threshold: Control how early eager processing begins relative to turn certainty.
• Automatic fallback: Reverts to standard turn handling if the transcript changes mid-turn.

Why it matters

• Reduces median response latency by ~150 ms when eager turns trigger.
• Improves tail latency, with top 5 percent savings reaching ~350 ms.
• Preserves conservative turn detection without added silence.
• Improves conversational flow without requiring configuration changes.

Example use cases

• Voice agents that avoid interrupting slower or hesitant speakers.
• Customer support flows with frequent mid-sentence pauses.
• Scheduling and intake agents handling long-form answers..

Getting started

1. Open your AI Assistant in the Mission Control Portal.
2. Navigate to Voice and scroll to the Transcription section.
3. Select Deepgram Flux as the transcription provider.
4. Set the Eager end-of-turn threshold based on how early you want tentative processing to begin.
   • The default value is 0.3.
5. Adjust the standard end-of-turn threshold as needed to control final turn confirmation.
6. Set both thresholds to the same value to disable eager behavior.