Everything K-culture — comebacks to K-beauty, straight to your inboxGet it in your inbox

METAL MEDIA

google-gemma/gemma-translator

468JavaScriptApache-2.0

A pocket-sized voice translator that runs Gemma entirely offline on a Raspberry Pi

Gemma Translator is an open-source kiosk device that translates spoken conversation between two people face-to-face, using an on-device AI model with no internet connection. It runs Google's gemma4-e2b model locally via LiteRT-LM, converts speech to text and back with Moonshine, and displays everything on a small retro-styled touchscreen. The whole thing is built to be assembled on a Raspberry Pi 5, including 3D-printable case files.

What it does

  1. Built with help from Google Antigravity as a two-lane 'face each other and talk' translator kiosk, where each side speaks their own language and hears the other's translation spoken back
  2. Runs the gemma4-e2b language model fully on-device via LiteRT-LM (Google's lightweight on-device inference runtime), so no cloud API calls are needed after setup
  3. Uses Moonshine for speech-to-text and moonshine-voice for text-to-speech, wrapped in a Python backend server and a React/Vite web frontend styled for tiny 480x320 kiosk screens
  4. Ships setup, model-download, start, and Raspberry Pi deployment scripts, plus a systemd service and 3D-printable STL case files for turning a Raspberry Pi 5 into a standalone kiosk appliance
  5. Supports two keyboard control schemes (single active-person mode with Space/Z/arrows, or two-hand mode with separate Z/X/arrow/plus-minus keys per person) for operating the translator without a touchscreen

Why it matters

It shows a concrete, working example of running a real conversational AI model completely offline on cheap, low-power hardware like a Raspberry Pi — useful for anyone building privacy-preserving or connectivity-free translation devices. The full open-source stack (model, speech pipeline, UI, hardware case) gives developers a ready template to fork for their own on-device AI kiosk projects.

Terms in this repo

  • LiteRT-LM · Google's runtime for running large language models efficiently directly on local devices, without a server
  • gemma4-e2b · A specific size/version of Google's Gemma language model designed to be run on-device
  • Moonshine · An open-source speech-to-text (and here also text-to-speech) system used to convert voice to text and back
  • push-to-talk · A control style where you hold a key/button down to record and release it to stop, common in walkie-talkies
  • systemd kiosk service · A Linux setup that automatically launches and locks a device into running one app in full-screen mode on boot

Repository description (English)

google-gemma/gemma-translator

Owner · google-deepmind

Open on GitHub

Coverage

Trending repos

All repos →

Latest from METAL MEDIA