Deepgram Flux

Deepgram launched Flux on 2 October 2025 under a label of its own making — conversational speech recognition. The distinction is real. Classical transcription models are built to turn a recording into text as accurately as possible; a voice agent needs something else, namely to know the moment the human has stopped talking and expects an answer. That decision is normally bolted on afterwards with a crude silence timer, which is why so many voice assistants either interrupt people mid-sentence or leave an awkward pause. Flux folds turn detection into the recognition model itself and reaches an end-of-turn verdict in under 400 milliseconds, while handling interruptions natively. On 29 April 2026 Deepgram made Flux Multilingual generally available, covering ten languages in a single model and — unusually — detecting and switching between them inside one conversation, without the caller having to declare a language up front. The model runs through Deepgram's cloud API with European endpoints, or self-hosted for organisations that cannot let audio leave their own infrastructure.

▸Videos

No videos yet.

▸Reviews