Google's Gemma Translator Runs Offline Voice Translation on an $80 Computer
Google's Creative Lab open-sources a fully offline, Raspberry Pi-powered voice translator built with Gemma 4, Moonshine ASR, and 3D-printed hardware
- What: Google Creative Lab open-sourced Gemma Translator, a fully offline, DIY voice translation device.
- Hardware: Raspberry Pi 5 (8GB RAM), portable battery, touchscreen, speaker, mini keyboard -- all held in a custom 3D-printed case.
- AI stack: Gemma 4 E2B via LiteRT-LM for translation, Moonshine for speech recognition, Kokoro for text-to-speech -- all on-device.
- Performance: Gemma 4 E2B runs at ~6 tokens/sec on the Pi 5, under 1.5GB RAM; Moonshine is up to 5x faster than Whisper Tiny.
- Deployment: One-command Pi setup via
deploy-pi.sh; also runs on laptops and Android devices. - License: Apache 2.0, free to clone and build -- STL files for the case are included in the repo.
Google's Creative Lab just dropped something that feels like a prop from a sci-fi film: a handheld, completely offline voice translation device you can build yourself. Gemma Translator is an open-source project that combines a Raspberry Pi 5, off-the-shelf components, and a stack of on-device AI models to translate speech between languages -- no internet, no API keys, no cloud bill.
The project is available right now on GitHub, free to clone and build. The hardware bill of materials is intentionally approachable: a Raspberry Pi 5 with 8GB RAM, a portable battery pack, a small touchscreen display, a mini speaker, and a micro keyboard with a rotary knob for language selection. A custom 3D-printed shell holds it all together, and STL files are included in the repo.
Three models, zero cloud calls
What makes this project technically interesting is that the entire pipeline -- speech recognition, translation, and text-to-speech -- runs locally. There are three distinct AI components working together:
- Translation: LiteRT-LM, Google's open-source edge inference framework, runs the
gemma4-e2bmodel. This is Gemma 4's smallest variant, a 2-billion effective parameter model that uses a mixed 2-bit/4-bit/8-bit quantization scheme to fit in under 1.5GB of RAM. - Speech recognition (ASR): Moonshine, an open-source ASR model from Useful Sensors, handles transcription. It's purpose-built for edge hardware and benchmarks at up to 5x faster than Whisper Tiny with no increase in word error rate.