An open-source handheld translator performs speech recognition, translation, and text-to-speech processing locally without requiring an internet connection.
The Gemma Translator is a handheld voice translation device that performs speech recognition, translation, and text-to-speech processing entirely offline. The open-source device is built around a Raspberry Pi 5 with 8GB of RAM and features a 480×320 touchscreen, microphone, speaker, push-to-talk buttons, rotary encoder, and a 3D-printed enclosure.
Google’s Gemma 4 E2B model has been deployed locally using LiteRT-LM for translation. Speech recognition is handled by Moonshine, while voice synthesis uses Kokoro. As a result, the entire translation process occurs locally without any need for cloud computing after deployment.
Two conversation modes are provided for one-to-one and face-to-face translation. Push-to-talk buttons and rotary encoders provide physical controls instead of relying solely on the touch-screen.
The translation speed is around 6 tokens per second while consuming less than 1.5 GB of memory on the Raspberry Pi 5. The project also provides STL files for building the enclosure and uses open-source software.
With speech recognition, translation, and synthesis performed locally, along with downloadable STL files for the enclosure, the project provides an offline approach to building a handheld voice translator.
