Skip to main content
​
April 8, 2026
FeatureUpdate
API v0.19.0

​
Graceful WebSocket session close

All WebSocket endpoints now support a close message for cleanly shutting down sessions. Send { "type": "close" } to signal you’re done — the server will finish processing any remaining audio and then close the connection. This replaces the previous cancel behavior and applies to both TTS and STT streams, including bridges.See the WebSocket protocol reference for full details.

​
Keepalive for STT streams

You can now send { "type": "keepalive" } on STT WebSocket connections to prevent idle timeouts during pauses in audio. This is useful for voice agent scenarios where the user may go silent but the session should stay open.

​
Endpointing controls for speech-to-text

Two new parameters are available for Deepgram Nova STT models:
  • endpointing — time in milliseconds of silence before the server finalizes speech. Set to false to disable automatic endpointing entirely. Defaults to 10.
  • utterance_end_ms — time in milliseconds of silence between words before sending an UtteranceEnd event. Range: 200–5000ms, default 1000ms.
These give you fine-grained control over how aggressively the API segments speech, which is especially helpful for voice agents and real-time transcription. See the Speech-to-Text API reference for supported models.

​
ElevenLabs TTS models available

ElevenLabs voice models — including Eleven v3, Flash v2, and Flash v2.5 — are now accessible through the SLNG API. Use the same unified endpoint format to synthesize speech with ElevenLabs voices alongside all other providers.

​
India region for Nova 3 Hindi

Deepgram Nova 3 Hindi is now available in the ap-south-1 (Mumbai) region, in addition to the existing ap-southeast-2 (Sydney) region. Use the X-Region-Override header to route requests to the closest region. See models by region for the full list.