TranscriptNet: a hybrid deep learning approach for digitizing and translating ancient Sanskrit manuscripts

Optical character recognition (OCR) allows the automatic transfer of handwritten or printed text into machine-readable format, whereas machine translation is aimed at transferring the digitised text of one language into another, which makes up an important pipeline towards the multilingual textual heritage. The linguistically rich Sanskrit language is particularly challenging, both in terms of deciphering correct character reconstructions and translation, due to its complex grammatical structure, diverse scripts, and limited available corpora. Ancient scripts are an invaluable part of our cultural heritage, and research remains inaccessible because conventional OCR and translation systems struggle with the script’s morphological complexity, degraded manuscript quality, and semantic nuance. Earlier studies have addressed either text extraction from manuscript images or basic text translation, but few have tackled the combined pipeline, especially for complex, real-world Sanskrit grammar. This study proposes a solution to the two-fold challenge of computational extraction and translation of ancient Sanskrit documents by adopting a two-stage deep learning pipeline, which combines convolutional and LSTM neural networks for character recognition and transformer-based neural language models for language translation. Key innovations include designing a large, representative character image dataset and developing advanced preprocessing algorithms to handle noise, segmentation, and script variability. The system uses three distinct datasets—Sanskrit character images, Nepali Handwritten Character dataset for handwriting variations and Bhagavad Gita sentence pairs to ensure linguistic and visual diversity during training. Additionally, translation models were trained and fine-tuned on corpora rich in grammatical complexity. Experimental results have shown that the proposed OCR model achieved a character error rate (CER) of approximately 2.5% and a word error rate (WER) of approximately 7.5%. The proposed pipeline attains an accuracy of 93% on the Nepali Handwritten Character dataset. The ability to digitise and translate ancient Indian texts written in Sanskrit allows for contributions to the areas of digital humanities, historical linguistics, and cultural preservation, offering a foundation for future research in multilingual archival access and AI-driven knowledge recovery from low-resource scripts.