// the find
niedev/RTranslator
Open source real-time translation app for Android that runs locally
RTranslator is an Android app that does real-time speech translation entirely on-device using Meta's NLLB for translation and OpenAI's Whisper for speech recognition. The main mode connects two phones over Bluetooth LE so each person hears the other speaking their own language through headphones. It's for anyone who needs offline, private translation — travelers, field workers, situations where cloud APIs aren't acceptable.
The on-device model optimization work is genuinely impressive: partial int8 quantization with KV cache and split model parts gets NLLB down to 1.3GB RAM with 4x inference speedup over a naive onnx conversion. The Bluetooth LE peer-to-peer architecture means no server, no account, no data leaving the device — the privacy story is real, not marketing. WalkieTalkie single-phone mode is a practical fallback that covers the most common tourist use case. The 3.0 roadmap replacing NLLB with HY-MT 1.5 and dropping MLKit is the right call — NLLB-600M translation quality has real gaps, and MLKit is the one closed-source piece currently.
6GB RAM minimum is a hard wall that cuts out a huge portion of Android devices globally, especially in the markets where language barriers matter most. The Bluetooth connection dropping bug has been listed as a known issue with no apparent fix timeline — for the primary use case, an unreliable connection during a real conversation is a serious problem. TTS quality is entirely punted to whatever system TTS the user has installed, which varies wildly and isn't something the app can control or improve. The model download (1.2GB) happens on first launch from GitHub, which means the onboarding experience fails entirely in regions with GitHub connectivity issues — sideloading workaround exists but most users won't find it.