Gemma Translator Brings Completely Offline AI Translation to Raspberry Pi 5

On-device AI is advancing very rapidly. It may have been just a few years ago when people were impressed by a microcontroller that could recognize a few basic hand gestures using an accelerometer and AI, but that’s child’s play today. An interesting open source project recently released by a small team at Google shows just how far the field has come. They built what they call Gemma Translator, a handheld device that packs a complete offline speech translation system into a custom Raspberry Pi 5-powered enclosure, allowing two people speaking different languages to have a conversation without sending anything to the cloud.

Unlike most translation apps that depend on remote servers, Gemma Translator runs the entire processing pipeline locally. Audio is captured through a microphone, transcribed, translated using Google's Gemma 4 E2B language model running under LiteRT-LM, and spoken back through a built-in speaker using Moonshine text-to-speech. Once the software has been installed, no internet connection is required, making the device useful in places without cellular service or Wi-Fi, or anywhere privacy is a concern.

The translator is built around a Raspberry Pi 5 with 8 GB of RAM and includes a small 480×320-pixel touchscreen, microphone, speaker, and a custom control panel with mechanical push-to-talk buttons and a rotary encoder for selecting languages. Everything is housed inside a bright yellow and black 3D-printed enclosure.

To drive the user interface, the developers built a lightweight software stack consisting of a React frontend optimized for small displays and a Python API server that communicates with the local language model. A single startup script launches the inference server, backend, and user interface together, and an automated deployment script can configure the system as a dedicated Raspberry Pi kiosk appliance.

The device’s interface was designed specifically for face-to-face conversations. Two keyboard modes allow either a single person to control both sides of the translation or two people to independently operate their own microphones and language selection. Physical controls handle recording and language changes.

To demonstrate the system, the team carried on a live conversation between Spanish and Mandarin speakers. The device transcribed each sentence, translated it locally using Gemma 4, displayed the translated text on the screen, and spoke the response aloud — all without contacting an external server.

Google has released the complete project, including the source code, setup scripts, deployment tools, and STL files for 3D printing the enclosure. With these resources and little more than a Raspberry Pi 5 and some basic components, you can make your own Gemma Translator.