Deepgram Flux

Deepgram Flux is the world's first conversational speech recognition model built specifically for real-time voice agents, combining STT and end-of-turn detection in a single neural model. Launched October 2, 2025, Flux eliminates the separate voice activity detection layer that traditional voice agent pipelines require, reducing false interruptions by approximately 30 percent and cutting agent response time by 200 to 600 milliseconds compared to STT plus VAD cascade architectures. Its semantic turn detection understands context rather than silence alone, enabling an optional eager end-of-turn event 150 to 250ms before the standard signal for latency-sensitive applications. Flux is accessed via Deepgram's streaming API at the /v2/listen endpoint with native integrations into Jambonz, Vapi, LiveKit, Pipecat, Cloudflare Workers AI, and Telnyx. Developers supply their own LLM and TTS components, keeping Flux as the purpose-built STT layer priced at $0.0077 per minute for English pay-as-you-go. Flux STT Multilingual, launched April 29, 2026, extends support to 10 languages including English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch, with native code-switching and mid-conversation language reconfiguration without reconnecting. Deepgram Flux is ideal for voice AI startup founders building outbound calling pipelines or low-latency conversational agents who already have a preferred LLM and TTS and need only a best-in-class STT plus turn detection layer. At $0.0077 per minute for English, it is one of the most cost-effective STT components for high-volume voice workloads, with Telnyx edge deployment removing public cloud hops for further latency reduction.