A Real-time Multi-lingual Android Translator Integrating Text and Voice Recognition

International audience In an increasingly interconnected global environment, the need for seamless multilingual communication via mobile devices is growing. This paper presents the design, development, and evaluation of a real-time multi-lingual Android system application translator (RT-MLAST), which integrates both text and voice recognition capabilities to convert spoken or typed input in one language to spoken and/or written output in a target language. The system pipeline comprises automatic speech recognition (ASR), neural machine translation (NMT), and text-to-speech (TTS) modules, augmented by language detection and multilingual support features. The study adopted a design science research (DSR) approach, which focuses on the creation and evaluation of an innovative artefact to solve a real-world problem. The ASR component was implemented using TensorFlow Lite to ensure compatibility with Android mobile devices. To enable efficient on-device operation, the NMT model was compressed using quantisation and pruning techniques. In order to collect data, a questionnaire was set up to check users' perception of the designed application. The Android application was implemented following the Model-View-Controller (MVC) design pattern to ensure modularity and scalability. The contributions of this work include a mobile-device real-time translator architecture, empirical measurements in the Android domain, and insights for deploying multilingual systems on resource-constrained devices. The results showed that the application achieved a whopping result of 99% accuracy when it was used for translation from one language to another using the intended means (Voice, text or image). The overall user acceptability based on the aspects was 3.50 out of 5. This suggested that users generally find the app quite acceptable, with only a few areas that could be improved (e.g., translation accuracy and responsiveness). For the Yoruba language, the application exhibited high accuracy in translating simple and direct sentences. However, it faced challenges when handling complex expressions and idiomatic phrases unique to Yoruba culture, which require a deeper understanding of linguistic nuances. The results show that the proposed system achieved competitive translation accuracy and high usability. The discussion implications for global communication, tourism, and education, and proposes directions for future work.