Whisplay HAT+ Rebuilds Google’s Offline Gemma Translator Into One Compact Pi 5 Slab

Google Creative Lab published Gemma Translator back in August 2026 as a voice box that never phones home. Alan Yam, Shashwath Santosh, and Dan Motzenbecker built it with help from Google Antigravity, then posted the code, model scripts, and 3D-print files on GitHub under Apache 2.0. You talk into a microphone. Moonshine turns the recording into text. Gemma 4 E2B, running locally through LiteRT-LM, translates that text. Moonshine speaks the other language from a speaker. After the first download, the radio can stay off.

Jadaie Lin modified the same setup to work on PiSugar’s Whisplay HAT+. The board is roughly the same size as a Raspberry Pi 5, measuring around 2.8 inches, but it also has some serious hardware, such as a 640 x 480 DSI touchscreen, two microphones and speakers, and a few motion sensors. The original Google setup required a separate kiosk panel (480 by 320), a USB microphone, a speaker, a push-to-talk button, and a rotary knob to select the language to use. Lin completed the entire process on a Pi 5 with only one hat put on top of it.

Audio turns up on the onboard mics, Moonshine does its magic with some sophisticated transcription, Gemma handles the translation, and the twin speakers return the outcome to you. There’s a small retro-terminal-style web UI in React and Vite that is fed via a Python API on port 3000. This displays two discussion lanes, allowing you to sit down in front of the tablet with a friend in separate languages and chat.

Lin relocated all of the old controls on the touchscreen and placed the language switching in a convenient settings panel. He also included timing to demonstrate how long each portion of the chat takes: listen, translate, and then speak.

Gemma 4 E2B runs in LiteRT-LM and lives in memory, with a useful system prompt stored in the key-value cache. Tokens are transferred to the browser via server-sent events, and the UI refreshes in real time as the discussion progresses. Finished phrases are fed into the speech queue. Context resets after 3200 tokens, as there is a 4096 token limit, thus extended conversations do not simply stop working. By default, the translator loads models for Chinese and English.

To set everything up, you’ll need Python 3.10 or newer, Node 18, or newer, an 8GB Pi 5, and a one-time model fetch from Hugging Face. After that, the device will happily run without the radio, which is ideal for travel, classrooms, and other situations where you don’t want your speech to sit on someone else’s servers. Lin is treating the translation as the first app for Whisplay HAT+, and there is a wait list and a Discord channel established up while the board is still being distributed to clients. The code is available at on the Github page, and it’s basically a fork of google-gemma/gemma-translator, so you can see how everything works.