Deepgram

Deepgram — AI Voice

Changelogs

Deepgram Flux adds Configure control message for mid-stream speech recognition updates//Deepgram's Flux speech-to-text API now supports dynamic configuration changes without disconnecting, allowing voice agents to inject context-specific vocabulary, adjust turn detection sensitivity, and modify timeout behavior on the fly. The feature is currently available via WebSocket connections with SDK support coming soon.

featureapisdk

Deepgram Voice Agent adds UpdateThink for mid-conversation LLM provider switching//Deepgram has released UpdateThink, a new client message that allows developers to dynamically replace the entire Think provider configuration during live Voice Agent conversations. This enables switching LLM providers, models, prompts, and functions in real-time without interrupting the conversation.

featureapiplatform

Deepgram releases self-hosted February 2026 container images with engine v3.107.0//Deepgram has released updated container images for its self-hosted speech recognition platform, including engine version 3.107.0 and updated API, license proxy, and billing components. The release requires NVIDIA driver version 570.172.08 or higher.

releaseplatform

Deepgram triples default concurrency limits for streaming speech and text APIs//Deepgram has increased default concurrency limits by up to 3X across Streaming Speech to Text, Text to Speech, and Voice Agent services. The improvements apply to Pay as you Go, Growth, and Enterprise plans.

platformfeatureapi

Deepgram adds OpenAI GPT-5.2 and Google Gemini 3 support to Voice Agent API//Deepgram's Voice Agent API now supports three new LLM models: OpenAI's GPT-5.2 Instant and Thinking variants, plus Google's Gemini 3 Flash. Developers can configure these models in their Voice Agent settings via a simple provider configuration.

featureapimodelintegration

Deepgram updates Nova-3 multilingual model with improved code-switching accuracy//Deepgram has released an updated Nova-3 multilingual speech-to-text model with lower word error rates across supported languages and significantly improved handling of code-switching scenarios. The improvements are available immediately without requiring API or configuration changes.

modelfeatureperformance

Deepgram adds Hebrew, Farsi, and Urdu support to Nova-3 speech-to-text model//Deepgram's Nova-3 speech-to-text model now supports three additional languages: Hebrew, Farsi (Persian), and Urdu. This expansion enables developers to build voice applications for these language communities with industry-leading accuracy.

featuremodelapi