Google Launches Gemini 3.5 Live Translate: Real-Time Speech-to-Speech in 70+ Languages, Translated as You Speak
The next cross-border meeting you run might genuinely not need that interpreter sitting in the corner anymore.
Three Key Facts
On June 9, Google unveiled Gemini 3.5 Live Translate, a speech-to-speech real-time translation model supporting more than 70 languages and over 2,000 language combinations. It automatically detects the language as you speak and generates translated speech, with output lagging only a few seconds behind the speaker. Google stresses the system tries to preserve the speaker's intonation, pacing, and pitch — not the flat robot-reading-a-script tone.
It's shipping across multiple Google products at once. In Google Meet, supported languages jump from the previous English-centric 5 to more than 70, opening first in private preview for select Workspace business customers, with a broader rollout later this year. The Google Translate mobile app supports it on both Android and iOS, and developers can tap in through the Gemini Live API. Android also gets a new "listening mode" that streams translated audio straight through the earpiece, so you can follow a translation privately without headphones.
But there's a glaring blank: Google published no hard numbers. No benchmark scores, no per-language-pair performance data, no specific latency figures, no comparisons to rivals. Which means the "few seconds behind" and "preserves intonation" claims currently rest entirely on Google's own word.
WangDou's Take
"70+ languages, 2,000+ combinations" sounds like Google dropped the UN interpreter booth into your phone — but the exact figures it declined to share are precisely the ones that matter most: how many seconds of lag, which pairs it actually nails, and who's accountable when it botches a sentence. The hard part of live interpretation was never recognition or pronunciation — it's that in those "few seconds" you have to listen, guess the full meaning of a sentence, and assemble the translation all at once, which is brutal for word-order-flipping pairs like Chinese-English. Google left benchmarks, latency, and per-pair performance blank and led with the sensorily pleasant "preserves intonation" pitch — a sign it isn't sure about the real hard part either. The honest test: later this year, run a three-way Chinese-Japanese-German meeting in Meet and see whether it interprets smoothly or freezes on the sentence that counts.
Source: Slator · 9to5Google · MultiLingual