Fluid, natural voice translation with Gemini 3.5 Live Translate

Details

Google DeepMind's Gemini 3.5 Live Translate advances real-time spoken language translation by processing speech as it streams rather than waiting for the speaker to finish. The model automatically detects over 70 languages and generates translated speech that preserves the speaker's intonation, pacing, and pitch — maintaining only a few seconds of lag throughout a session. The rollout targets three audiences simultaneously: developers can access the model in public preview via the Gemini Live API and Google AI Studio; enterprises can test it in private preview via Google Meet starting this month; and consumer users will get it through Google Translate on Android and iOS. Developer platforms including Agora, Fishjam, LiveKit, Pipecat, and Vision Agents have already integrated with the API. Early real-world testing includes Grab, which handles over 10 million voice calls per month between drivers and riders and is trialing the model for multilingual communication at pickups. The model is described as robust to noise and handles multilingual inputs without manual language configuration. The practical implication for builders is direct: real-time speech translation is now accessible via API without managing separate encoder pipelines or manual language detection. The Gemini Cookbook includes demo code for the Gemini Live API integration.