A two-way offline translator that turns a Pi into a conversation kiosk

What it does

Gemma Translator is an on-device voice translation appliance built around a Raspberry Pi 5. It listens to speech through a microphone, transcribes it with Moonshine, translates via a local Gemma 4 model running in LiteRT-LM, and speaks the result back in the other language. The interface is split into two lanes—one for each speaker—with retro-terminal styling meant for small 480×320 displays. A unified launcher script handles the local LLM server, Python API backend, and React frontend together.

The interesting bit

The whole thing is designed as a physical appliance: the repo includes STL files to 3D-print the case, and a deployment script that locks the Pi into a systemd kiosk running Chromium in full-screen. The two-lane UI and dual keyboard modes—one-handed “active person” or two-handed “vertical”—treat the device like a shared arcade cabinet for conversation.

Key highlights

• Runs the gemma4-e2b model entirely on-device via LiteRT-LM with no network after setup.
• Retro-terminal React frontend optimized for tiny handheld or touchscreen displays.
• Includes STL files and a one-command Pi deployment script that sets up a systemd kiosk service.
• Two keyboard-driven interaction modes for either single-operator or two-person use.
• Uses Moonshine for both speech-to-text and text-to-speech.

Caveats

• On-screen touch controls are not enabled; all interaction is keyboard-only.
• Hardware targets are specific: it expects a Raspberry Pi 5 with 8GB RAM, and setup scripts assume Linux or macOS.
• This is an experimental Google Creative Lab project, not an officially supported Google product.

Verdict

Hardware tinkerers building offline LLM appliances or low-cost translation kiosks will find a complete recipe here. If you need a general-purpose translation API or a mobile app, this is too tightly coupled to its Pi-and-keyboard shell.