Deepgram's Flux Multilingual puts turn detection in 10 languages

Deepgram's Flux Multilingual extends a voice agent's ability to judge when a caller has finished speaking across ten languages. The release article says transcription and turn detection operate together. Instead of treating a pause of a fixed length as the end of every reply, the model uses a learned confidence signal about whether the conversational turn is complete.

Developers can tune that signal, which the article describes as dialing toward faster responses or toward more conservative end-of-turn decisions depending on the use case. This gives applications a direct control over how readily the assistant takes its turn.