Deepgram Voice Agent API

Deepgram Voice Agent API wraps the Nova-3 speech recognition model, a managed LLM, and Deepgram Aura TTS behind a single WebSocket endpoint. The Flux conversational recognition model, which became generally available in October 2025, adds integrated end-of-turn detection with approximately 260ms decision latency, reducing silence-padding overhead common in earlier STT-based pipelines. STT accuracy on Nova-3 reaches 5.26% batch WER across a 36-language corpus, among the lower error rates available in managed voice agent cascades. Standard pricing is $0.075 per minute with an option to substitute a custom TTS provider for $0.065 per minute. The Growth plan supports up to 60 concurrent sessions in North America. Official SDKs in Python, JavaScript, C#, and Go allow integration across backend stacks, while native connectors for LiveKit, Pipecat, Vapi, Retell, and ElevenLabs reduce time to production. Inbound and outbound telephony integration is available natively through the platform. Deepgram Voice Agent API is the practical first choice for startup developers building production voice agents who need verified STT accuracy, broad SDK coverage, and a straightforward integration path. The hosted documentation, example repositories, and active developer community lower onboarding friction, while the BYO TTS option gives cost-conscious teams an early optimization path as call volumes scale.