compareeHub

GOOGLE'S FREE DESIGN FOR A TRANSLATOR THAT NEEDS NO INTERNET

A pocket box that translates a conversation both ways, entirely on-device, with no internet after setup.

by Google Creative Lab (Alan Yam, Shashwath Santosh, Dan Motzenbecker)

FULL CAD BOM FIRMWARE DOCS

AIOpen-hardware

Built withRaspberry Pi3D printing

difficulty
●●●○○
time
a weekend
license
Apache-2.0
repo
repo ACTIVE1,629 stars

WHAT YOU’LL NEED

Videos
1
Jump to section

COMPAREE VERDICT

Gemma Translator is an offline two-way voice translator from a small team at Google Creative Lab. It runs on a Raspberry Pi 5 with 8GB, using Google's compact Gemma 4 E2B model through LiteRT-LM for translation, Moonshine for speech recognition and the moonshine-voice package for speech output, so after the one-time model download nothing needs the internet. The code offers six languages: Arabic, English, Spanish, Japanese, Chinese and Korean. According to the video, all parts are off the shelf: the Pi 5, a portable battery pack, a small touchscreen, a mini speaker and a micro keyboard with a knob for choosing languages, held together by a 3D printed case (three STL files in the repo). The controls are plain keyboard events: hold Z to talk, arrow keys or the knob to change language, Space to switch between the two people, so there is no GPIO wiring at all. The gap is the shopping list. The README names the part types but not exact models, so you choose the display, speaker, microphone and keyboard yourself and check they fit the case. The software side is clean: a few shell scripts, and deploy-pi.sh turns the Pi into a kiosk that starts on boot.

GOOD TO KNOW

  • —Apache-2.0 licence, commercial use allowed.
  • —README gives the setup scripts (setup.sh, download_model.sh, start.sh, deploy-pi.sh) and full run instructions.
  • —Three STL files in stl/ folder: front-face, main-body, speaker-support.
  • —No exact parts list: the README names part types (Pi 5 8GB, microphone, speaker, display), not models; controls are keyboard keys, no GPIO wiring.
  • —Code is Python (http.server API with Moonshine speech recognition and moonshine-voice speech output) and React (Vite frontend); the model is Gemma 4 E2B via LiteRT-LM.
  • —Google notes this is not an officially supported product and is not eligible for their vulnerability rewards programme.

Parts to buy

6 items

From our check of the build. Exact quantities and part numbers are in the creator’s BOM.

  • Raspberry Pi 5 8GBFind
  • Microphone or USB audio interfaceFind
  • SpeakerFind
  • Small display (for example 480x320)Find
  • Small keyboard or keypadFind
  • Video also uses a portable battery packFind

Can I build this?

PrintThree STL files (front-face.stl, main-body.stl, speaker-support.stl), print time not stated.
BuyRaspberry Pi 5 8GB, microphone or USB audio interface, speaker, small display (for example 480x320), a small keyboard or keypad; the video also uses a portable battery pack. Exact models are not documented.
Tools3D printer, a computer to prepare the Pi's SD card, Raspberry Pi OS, internet for the one-time setup. The scripts install Python packages and Node.js themselves.
SkillsIntermediate: comfortable setting up a Raspberry Pi, running shell scripts and picking parts that fit the printed case. No soldering or GPIO wiring needed.
TimeA weekend: printing the case, assembly and running the setup scripts.
Cost$$ — the Raspberry Pi 5 8GB is the biggest cost; the repo gives no prices.
SafetyNone beyond ordinary electronics care: low voltage only, no mains. If you add a battery pack, use a ready-made power bank rather than bare cells.

Build at your own risk. Projects involve tools, electronics and sometimes mains voltage — follow the creator’s safety notes.

Videos

Gemma Translator: offline DIY translation device built with Gemma 4 + Antigravity1:14

Official video from the Google Antigravity YouTube channel showing the finished device and how it is built, published 5 Aug 2026.

More builds like this

All projects
A generative image model running on a one-dollar microcontrollerAICrypto miners dumped this PS5-based chip, now it runs AI at homeAIA 28.9M parameter language model runs entirely on a cheap ESP32-S3 microcontrollerAI

Gallery

Start here

Navigation into the creator’s own docs — we don’t rewrite the guide, we route you to the source.

  1. 1.Read the README hardware section and pick your Pi 5 8GB, microphone, speaker, display and a small keyboard or keypad (The README names part types only. The video shows a Pi 5, battery pack, small touchscreen, mini speaker and a micro keyboard with a knob.)
  2. 2.Print the three STL files (front-face, main-body, speaker-support) from the stl/ folder (No print settings or print time given; test-fit your display and speaker before final assembly.)
  3. 3.Install Raspberry Pi OS, clone the repo, run setup.sh, then download_model.sh to fetch Gemma 4 E2B (The model download needs internet and about 6GB of free space.)
  4. 4.Connect the microphone, speaker, display and keyboard, then run deploy-pi.sh to install the kiosk service (The deploy script installs the packages, builds the UI, registers a systemd service and sets LXDE autostart to open Chromium in kiosk mode at localhost:3000.)

Resources

Documentation, files and community threads for this build — we link straight to the original sources and never rehost the creator’s files.

KNOWN ISSUES

  • No exact parts list. The README names the part types (Pi 5 8GB, microphone, speaker, display such as a 480x320 kiosk screen) and the video adds a battery pack and a micro keyboard with a knob, but no models or links, so check that your parts fit the printed case.
  • The README gives 480x320 only as an example display size and names no model or interface. Check that your display works with the Pi 5 and fits the front face before buying.
  • download_model.sh pulls the Gemma 4 E2B model (about 2.6GB) from Hugging Face and needs about 6GB free, so setup needs internet even though translation later runs offline.
  • deploy-pi.sh assumes Raspberry Pi OS / Debian with the LXDE desktop session for kiosk autostart; on other systems you set up the service and browser yourself.
  • Recording and language changes are keyboard-only in the current build and on-screen touch controls are not enabled, so you need a keyboard or a keypad that sends the right keys (Z, X, Space, arrows).
  • Google states this is not an officially supported product and is not eligible for its open source vulnerability rewards programme, so do not count on long-term maintenance.

Does this work on a Raspberry Pi 4 or a Pi with less than 8GB?

The README lists a Raspberry Pi 5 with 8GB of RAM as the compute board and the deploy script targets the Pi 5 (8GB). Smaller Pis are not covered by the project.

Which languages does it support?

The language selector in the code offers six languages: Arabic, English, Spanish, Japanese, Chinese and Korean. Speech recognition and speech output are set up for the same six.

Can I use this commercially?

The code is Apache-2.0, which permits commercial use. Google states this is not an officially supported product, so there is no warranty or support from Google.

Do I need internet after setup?

No. You need internet once to install packages and download the Gemma 4 E2B model (about 2.6GB, the script wants 6GB free). After that, translation, speech recognition and speech output all run on the Pi.

What is the translation accuracy?

The README gives no accuracy figures. Gemma 4 E2B is a compact on-device model, so do not expect the quality of large cloud translators, especially for complex or specialist language.

Community builds

No community builds yet — be the first, we feature the best ones.

Discussion1

FROM THE COMPAREE TEAM

Six languages, all running on a Raspberry Pi 5 with no internet. Which language pair would you want first, and would you trust it in a real conversation?

CompareeTEAM3d ago

Practical notes from our verification: the repository is under the google-gemma organization (last push October 2026), the video is live on the Google Antigravity channel, and the three STL files are in the stl/ folder. The biggest gap is the shopping list: the README names part types (Pi 5 8GB, microphone, speaker, display), not models, so you pick parts yourself and check they fit the case. The controls are ordinary keyboard keys, so a small keyboard or keypad is all the input hardware you need; there is no GPIO wiring. The software deployment is excellent: a few shell scripts and you are running. Google states this is not an officially supported product.

Google Creative Lab (Alan Yam, Shashwath Santosh, Dan Motzenbecker)

Gemma Translator was made by a small team at Google Creative Lab (Alan Yam, Shashwath Santosh and Dan Motzenbecker). The repo was built with the assistance of Google Antigravity and shows an offline voice translator running Google's Gemma 4 model through LiteRT-LM on a Raspberry Pi 5.

GitHub Web

Star the project on GitHub

DISCLAIMER

  • Comparee is not the author of the projects featured here. All rights to each project belong to its creator — every page links to the original source, and we never host creators’ files.
  • Information is provided without warranty and may become outdated as projects evolve. Prices are indicative bands only — always check the creator’s parts list for current costs.
  • Building and operating any project is at your own responsibility. Protective equipment, safe workshop practice and compliance with local regulations are the builder’s responsibility.
GitHub repo