Google just dropped something I’ve been waiting years for.
Gemini 3.5 Live Translate is out today. Real-time, speech-to-speech translation in over 70 languages. And I don’t mean a demo or a staged keynote bit. It’s live. Right now. In the Translate app, the Gemini API, AI Studio, and coming to Google Meet.
What It Actually Does
This isn’t your standard Google Translate where you tap a language pair and hold your phone up to a sign. Gemini 3.5 Live Translate is a speech-to-speech model.
It listens in one language and speaks back in another. It preserves your intonation, pacing, and pitch. So the translated version sounds more like you, not some generic TTS robot.
The model generates speech continuously instead of waiting for you to stop talking. That’s the key technical difference. It stays just a few seconds behind throughout the conversation. No awkward dead air while waiting for the translation to catch up.
It’s also noise robust, so it works in loud environments. Google wraps every output in SynthID watermarks so there’s always a way to detect AI-generated audio.
Where You Can Use It
The rollout is surprisingly broad for a day-one launch.
Gemini Live API and AI Studio. Developers can start building with it today via the Gemini Live API. Logan Kilpatrick confirmed the model is available in both the API and AI Studio right now.
Google Translate app. This is the big one for most people. The model is rolling out globally on both Android and iOS through the Live Translate feature. Any pair of headphones works. You no longer need Pixel Buds or any specific hardware.
Listening mode (Android). This is clever. If you don’t have headphones, you can hold your phone to your ear like a regular call and hear the translation through the earpiece. I can see this being huge for museum tours, conferences, or just quick bike-lane conversations.
Google Meet. Enterprise customers get a private preview starting this month. The upgrade is significant: from 5 supported languages to 70-plus, and from English-only translation to any-to-any across more than 2,000 language combinations.
The Backstory
Google has been working on real-time translation for a long time. The company has been carefully vague about it. They’ve previously argued it isn’t safe to AI generate someone’s voice without watermarking it or to rely on models making errors.
The Translate app got Gemini-based live translation late last year for some users, but it was limited. This is the full version built on Gemini 3.5, the model family Google launched at I/O 2026. I covered the previous Gemini generation when it rolled out as the default across Google’s products.
Google Translate itself just passed 20 years. The service now supports nearly 250 languages, handles over a trillion words per month, and serves more than a billion users. This is the natural next step for that pipeline. It’s also happening alongside Apple leaning on Gemini for their own AI push.
Partners are already testing it. Grab is using it for driver-rider translation. CJ ENM is looking at it for media dubbing. Platforms like LiveKit, Agora, and Pipecat have integrated the API so developers can skip the real-time media infrastructure work and just build.
The Bottom Line
I’ve been skeptical of real-time translation for years. The demos always look great and fail in the real world. But this feels different.
Three things stand out. One, it’s available on day one across multiple products, not locked to one device. Two, the continuous generation approach is architecturally real, not a marketing trick. Three, Google watermarked everything with SynthID before shipping instead of promising to add safety later.
The usual shipping gap story on this site is that indie builders ship while trillion-dollar companies demo, stall, and get sued. But sometimes the big company actually delivers. Today was one of those days.
Try it. Open Google Translate, grab some headphones, and talk to someone in a language you don’t speak. That’s a real thing you can do right now.
