google-gemma/gemma-translator

Gemma Translator is a project for building a physical, standalone voice translation device using a Raspberry Pi. It was built with help from Google Antigravity and runs entirely offline once set up, using Google's Gemma 4 language model through a tool called LiteRT-LM to do the actual translation locally on the device, with no internet connection needed afterward. Speech is converted to text and back to speech using a system called Moonshine. The intended hardware is a Raspberry Pi 5 with 8 gigabytes of memory, a microphone, a speaker, and a small screen such as a 480 by 320 pixel display, the kind commonly used in handheld or kiosk style projects. The software includes a web based frontend built with React, styled to look like a retro computer terminal and sized for small screens, along with a Python server that connects the frontend to the translation model. The design supports two people speaking different languages at once, described as two lanes, each with its own rotating list of languages. There are two keyboard modes for controlling it: a single operator mode where one person switches which lane is active and controls everything from one set of keys, and a two person mode where each side has its own dedicated keys so two people can each manage their own language and recording independently. Setup involves running provided scripts that create a Python environment, install dependencies, and download the Gemma model. A separate deployment script automates turning a Raspberry Pi into a dedicated kiosk device, including setting it up to start automatically and display the translator full screen when the device powers on.