Gemma Translator — offline speech translation on a Raspberry Pi 5

Google ran a live Spanish to Mandarin conversation on a Raspberry Pi with no internet access. Meet Gemma Translator, which handles the translation on the board itself. Moonshine does the speech recognition, and a third model generates the voice reading it back. All three run on the Pi, no API calls, no server, no network. It pushes about seven tokens a second, which is slow for a language model and fast enough to hold a conversation. And that's where the hardware stops being the limit. A Raspberry Pi 5 costs less than a pair of headphones. Add a battery pack, touch screen, a speaker, and a micro keyboard with a knob for switching languages. Two modes let one person drive both sides or two people run their own mics. Print the case yourself. The STL files are included. Here's where it gets insane. Google published all of it. The code, the parts list, the deployment script, the enclosure files. You can clone the repo and build one. But the craziest part? Running a model like this used to mean a server rack and a connection to it. This one fits in your hand. If you want to stay updated with the latest tech news, follow for more.