Google is taking the awkwardness out of real-time translation. The company just launched Gemini 3.5 Live Translate, a speech-to-speech model that can handle over 70 languages with near-zero latency. Unlike traditional turn-based translation that waits for a full sentence, this model streams continuously — it balances gathering enough context for accuracy with spitting out audio quickly enough to avoid those painfully long pauses. The result? Conversational lag measured in seconds, not awkward silences.
The model is also impressively noise-resistant and can detect mixed-language input without any configuration. Even better, it preserves the original speaker’s tone, rhythm, and pitch, so you don’t sound like a robot. Several real-time audio/video frameworks — including Agora, Fishjam, LiveKit, Pipecat, and Vision Agents — have already integrated with the Gemini Live API, saving developers the trouble of building their own streaming infrastructure. On the application side, ride-hailing giant Grab is testing the model in over 10 million calls per month to help drivers and passengers communicate across languages. South Korean entertainment conglomerate CJ ENM is also experimenting with the tech for live interpretation and dubbing for global audiences.
As for availability, developers can access the public preview through Google AI Studio and the Gemini Live API. Starting this month, enterprise users get a private preview in Google Meet, with language support jumping from 5 to over 70 languages — that’s more than 2,000 possible language pairs in a single meeting, no longer requiring English as an intermediary. For individuals, the Google Translate app on Android and iOS now taps into the same capability. Not bad for a service that started as an experimental project back in 2006 — today, Google Translate has over 1 billion monthly active users.