I want to say this plainly before I say anything else: Google’s engineers built Gemini Live Translate. Google published the starter code for it. I did not invent real-time speech translation. What I actually did is narrower, and honestly more interesting to write about: I took a script Google gives away for free and turned it into a service I could open from my own phone, on my own domain, before I had confirmed Google’s own consumer app was reliably doing the same thing for me here in Korea.
What “Get Code” Actually Gives You
Google’s AI Studio lets you try a model in a browser playground first. Once you like what you see, there’s a “Get Code” button that hands you a runnable starter script, already wired up to the right model name and the right authentication pattern, so you’re not starting from raw API documentation.
For Gemini Live Translate, that starter script is a Python program from Google’s own public cookbook repository. Stripped down to what actually matters, it does three things: opens a live connection to models/gemini-3.5-live-translate-preview, sets a translationConfig with a target language code, and streams microphone audio in while streaming translated audio back out. The rest of the script is desktop plumbing, PyAudio for the microphone, OpenCV for an optional camera feed, a screen-capture option using mss. None of that plumbing is what I actually wanted.
What I Actually Wanted
On July 13, I asked my AI coding assistant whether this newly announced technology could be applied to an app I already had, the older Flutter translator sitting in Play Store internal testing. The first step wasn’t a rewrite. It was verification: get the starter script running, confirm the translation actually worked the way Google described it, on my machine, with my own key.
It did. And once it did, the desktop Python script stopped being the point. I didn’t want a script I ran from a terminal with a webcam attached. I wanted a page anyone in my house could open on a phone, served from my own NAS, reachable through my own subdomain, with no port forwarding and no server holding anyone’s audio. So the actual rebuild was a translation in a different sense: the same wire protocol Google’s script speaks, model name and translationConfig and all, rewritten from Python’s PyAudio and asyncio into the browser’s own Web Audio API and a raw WebSocket, so it could run entirely client-side on translee.boseonggang.com.
That move is where the real friction showed up, including the WebSocket 1007 error I cover in detail in the second post of this series, a config field that worked one way in Google’s own sample and needed a different nesting to work from a browser. Google’s starter code gets you a working proof of concept. It does not tell you how to turn that proof of concept into a deployable app with your own domain, your own PWA install prompt, and your own decisions about where an API key lives. That gap, between “the demo works” and “this is a real thing I can hand my family a link to,” is most of what the rest of this series is actually about.
The Claim I’m Not Making, and the One I Am
I looked into Gemini Live Translate’s actual rollout while writing this post, because I didn’t want to say something untrue for the sake of a good line. Here’s what I found. In December 2025, Google shipped an earlier wave of live translation improvements to the Google Translate app, but explicitly limited to the US, Mexico, and India, with other countries including Korea promised for later in 2026. Then on June 9, 2026, an entirely new model, Gemini 3.5 Live Translate, launched much more broadly at once: the Live API and AI Studio for developers, and the Google Translate app on Android and iOS, described as a worldwide release.
I haven’t independently confirmed whether every piece of that worldwide release, including features like an Android-only mode where you hold the phone to your ear and it speaks a translation directly, is actually live and working the same way for me in Korea today. I don’t want to overstate that gap. What I know for certain, because I watched it happen on my own machine, is this: I had a working version of this technology, streaming, auto-detecting language, speaking translations back, running on my own phone, before I went looking for Google’s official consumer app and checked whether it did the same.
That’s the honest version of the claim. Not “before Google.” Before I checked what Google’s own app could already do for me. AI Studio’s “Get Code” button is exactly what makes that possible for someone who isn’t a professional developer: a straight line from a demo Google is proud of, to a real app running on infrastructure I own, without waiting for a feature flag to reach my country first.
I still don’t have real cost numbers for the Gemini API usage behind this app. That’s an honest gap I’ll close in a future post once I’ve used it enough days to report something true instead of a guess.