Abstract: The rapid growth of digital communication has increased the demand for efficient and accurate language translation systems. However, most existing translation tools rely heavily on internet connectivity, limiting accessibility in low-network regions, and raising significant data privacy concerns. Furthermore, traditional rule-based translation systems often produce contextually incorrect and slow translations, particularly for less-resourced languages such as Nepali and Sinhala. To address these limitations, this paper presents an AI-Based Text-to-Text Machine Translation System that translates Nepali and Sinhala text into English in an offline environment. The proposed system leverages deep learning techniques using PyTorch and HuggingFace Transformers to implement a pre-trained neural machine translation model capable of generating accurate and context-aware translations. The system is developed using Python and Django, ensuring a secure, user-friendly interface while maintaining offline functionality. The architecture consists of multiple modules including input handling, text pre-processing, neural translation engine, post-processing, and result display. By eliminating internet dependency and ensuring local processing, the system enhances privacy, accessibility, and reliability. Experimental evaluation demonstrates that the proposed solution provides efficient, grammatically coherent, and contextually meaningful English translations while maintaining complete offline usability.

Keywords: Machine Translation, Low-Resource Languages, Nepali, Sinhalese, Transformer, Transfer Learning, Sub word Tokenization

Downloads: | DOI: 10.17148/IARJSET.2026.13360