Gemma Translator: offline DIY translation device built with Gemma

Gemma Translator: offline DIY translation device built with Gemma 4 + Antigravity

Meet Gemma Translator, a completely offline DIY translation device powered by Gemma 4, built with Antigravity.

All the parts are off the shelf and easy to find: a Raspberry Pi 5, portable battery pack, tiny touchscreen display, mini speaker, and micro keyboard with a knob to select languages. It’s all held together by custom 3D-printed housing.

We use LiteRT-LM to run Gemma (gemma4-e2b) on-device with a simple Python server to handle requests. All the translation knowledge is built right into the model, so no APIs or Wi-Fi are required. Both speech recognition and text-to-speech run on-device too, using Moonshine for ASR and Kokoro to turn Gemma’s output text into speech.

The UI is a simple web frontend in React with a monochrome theme that shows info like language selection and mic feedback, handling hardware controls like knob turns and push-to-talk as keyboard events.

You can build it all yourself by cloning our repo: https://github.com/google-gemma/gemma…

Gemma Skills – https://github.com/google-gemma/gemma…

Google Antigravity is an agent-first software development platform designed to help developers build, orchestrate, and monitor AI agents. It connects AI agents directly to code editors, terminals, and browsers so they can perform development tasks autonomously.

The Gemma Translator is a portable, offline device built using easily accessible, off-the-shelf components. To replicate this project, you will need the following hardware:

Raspberry Pi 5: The core computing unit powering the system (0:07).
Portable battery pack: Ensures the device can be taken anywhere for true offline use (1:04–1:06).
Tiny touchscreen display: Provides the visual interface for the translation device.
Mini speaker: Used for the text-to-speech output.
Micro keyboard: Includes a physical knob to select languages and controls for inputs.
Custom 3D-printed housing: Designed to hold all the electronic components together in a single unit (0:31–0:32).
Microphone: Essential for speech recognition (0:10–0:11).

Build Google's Gemma 4 Translator | Offline AI Translator | Open source

The Gemma Translator is an open-source, 100% offline DIY translation device built by Google Creative Lab using the local Gemma 4 E2B AI model and developed through the Google Antigravity platform. This portable handheld gadget processes speech recognition, live translation, and text-to-speech entirely on-device, requiring absolutely no internet or cloud connectivity. [1, 2, 3, 4]

The full blueprints, code, and 3D-printing assembly files have been published under an Apache 2.0 license on the official Gemma Translator GitHub Repository. [2, 3, 5]

Core Technology Stack

The engineering architecture leans heavily on optimized local execution models to bypass cloud processing entirely:

Hardware Framework: The device is developed using Google Antigravity and targets a Raspberry Pi 5 (8GB RAM recommended).
Translation Brain: It runs Google’s lightweight open model, Gemma 4 (gemma4-e2b), optimized for edge deployment using LiteRT-LM.
Hearing (ASR): Uses Moonshine for local, multi-language speech-to-text processing.
Speaking (TTS): Leverages Kokoro (or integrated Moonshine modules) to synthesize the translated text back into clear audio speech. [1, 2, 3, 4, 6]

Device Hardware & Features

The physical prototype is fully self-contained and resembles a dedicated translation radio: [2]

3D-Printed Housing: A custom enclosure designed to neatly fit all electronic components.
Display Interface: A small touchscreen displaying live, dual-language translated text in real time.
Manual Controls: Features a built-in push-to-talk button and a physical rotary dial to quickly select languages.
Audio Peripherals: Outfitted with a mini microphone and a standard speaker for seamless verbal interaction. [1, 2]

Key Benefits

Total Data Privacy: Sensitive conversations remain entirely local and never touch remote servers.
Remote Portability: Functions seamlessly in airplanes, remote rural regions, or areas with restricted or zero cellular coverage.
Open DIY Blueprint: Allows tech enthusiasts and hobbyists to build, iterate, and customize their own offline translation hardware using cheap off-the-shelf components. [1, 2, 7]