Gemini 3.5 Live Translate puts near real-time speech-to-speech into 70+ languages

Google introduced Gemini 3.5 Live Translate, a speech-to-speech translation model that runs on streaming audio instead of waiting for a speaker to finish. It auto-detects the input language, supports more than 70 languages and over 2,000 language pairs in a single meeting, and aims to stay only a few seconds behind the speaker. Output is watermarked with SynthID, and Google says the model is built to keep working through background noise and crosstalk rather than ideal-mic conditions.

The interesting choice is in the tradeoff. Older translation systems wait for a natural pause, then translate the whole utterance well; Live Translate generates target-language audio continuously, so latency drops at the cost of having to revise its translation mid-stream. Google is shipping it in three places first. Grab is testing it on its driver-passenger voice calls, which it puts at more than 10 million per month, so a rider in Jakarta can speak Bahasa Indonesia and the driver hears their own language. Google Meet is expanding from five translation languages to the full 70+, with arbitrary pairs instead of English-as-bridge. And on Android, a listening mode plays translated audio through the phone earpiece without headphones.

Why it matters

If you build voice products with global users, the bar moves: the default user expectation will shift from "captions in another language" to "they hear me in theirs, in conversation." For teams running multilingual support, sales, or rideshare-style flows, this is the first credible alternative to staffing native speakers on every shift; worth piloting on calls where current translation already breaks down on accents or background noise.