Real-Time on-Device Neural Machine Translation for Dynamic Multilingual Mobile Application Localization

MobileNMT, a small on-device neural machine translation solution, dynamically localises static UI and APIfetched data in Android apps without strings.xml localisation files. With a quantised encoder-decoder Transformer architecture optimised for TFLite, MobileNMT achieves a compact model size (12.6 MB) and low inference latency ($\sim 90$- 100 ms) on mid-range mobile hardware. Shared embeddings and parameter-efficient compression configurations are key developments. Based on comprehensive tests, MobileNMT achieves a BLEU score of 24.9, maintains runtime memory below 50 MB, reduces APK size by 70%, allows offline operation, and protects user privacy. Scalability and performance are improved over static localisation files and important translation baselines like Google ML Kit (offline) and uncompressed Transformers. MobileNMT appears to be a flexible and stable option for multilingual mobile app development in dynamic, resource-constrained contexts.