Build an Offline Gemma Translator on Raspberry Pi 5

Assemble and identify the exact hardware

Install the active cooler, insert the prepared microSD card, then connect the USB microphone, speaker, display, keyboard, network and 27 W power supply.

• Confirm the board has 8 GB RAM.
• Do not hide the active cooler under an incompatible case.
• Keep Ethernet or Wi-Fi connected until every model and language asset has been downloaded.

Install Raspberry Pi OS and the host packages

Flash 64-bit Raspberry Pi OS Bookworm, boot to the desktop, open a terminal and update the system before installing the build, audio and service tools used by the tutorial.

Raspberry Pi Imager →

sudo apt update && sudo apt full-upgrade -y
sudo apt install -y git python3-venv python3-pip ffmpeg libasound2-dev pulseaudio-utils alsa-utils lsof netcat-openbsd nodejs npm

Switch the kiosk session from Wayland to X11

The reviewed deployment script writes an LXDE X11 kiosk autostart entry. Select X11 for the tutorial path, reboot and return to the terminal.

sudo raspi-config nonint do_wayland W1
sudo reboot

Prove microphone and speaker operation first

List the capture and playback devices, then make a short recording and play it back. Do not start the AI installation until this simple audio loop works.

arecord -l
aplay -l
pw-record test.wav
pw-play test.wav
rm test.wav

Note: Press Ctrl+C to stop pw-record before running pw-play. An HDMI monitor may appear as an additional audio output.

Clone and pin the reviewed Google project

Pin the source revision reviewed by LocalClaw so that a later upstream change cannot silently alter this guide.

git clone https://github.com/google-gemma/gemma-translator.git
cd gemma-translator
git checkout 47f9b3ba40ca3650fb80ee42264a76d6a2b5f8ba
chmod +x setup.sh download_model.sh start.sh deploy-pi.sh

Apply the documented dependency-script fix

At the reviewed commit, setup.sh enables pip hash enforcement while requirements.txt contains no hashes. Remove only that incompatible flag before deployment.

grep -- --require-hashes setup.sh
sed -i 's/--require-hashes //' setup.sh

Note: This fix is specific to the reviewed source state. Inspect setup.sh first and skip the edit if upstream has already removed the flag.

Run the one-command appliance deployment

The script installs the Python environment, builds the React frontend, imports the 2.6 GB LiteRT-LM model, creates the systemd service and configures Chromium kiosk mode.

./deploy-pi.sh

Note: DroneBot Workshop observed roughly 20–40 minutes for this stage. Keep at least 6 GB free and do not interrupt the model import.

Reboot, wait for the service and test a translation

After reboot, Chromium should open the local interface. Select the source and target languages, hold Z while speaking, then release it to transcribe, translate and play the result.

sudo reboot
systemctl status --no-pager gemma-translator.service

Note: The kiosk can appear before the backend is ready. Wait several seconds before treating an initial localhost error as a failure.

Cache language assets and prove the offline boundary

Use every language pair you plan to rely on while networking is still available. Then disable Wi-Fi or unplug Ethernet and repeat a translation.

• A translation after disconnection proves local inference only for assets already downloaded.
• Record which language pairs were tested.
• Do not claim all languages are offline until each required Moonshine asset has been cached.

Note: DroneBot Workshop demonstrated English speech translated to French output with Ethernet unplugged and no Wi-Fi configured.