IIT Roorkee Unveils World’s First AI to Transliterate Modi Script
Share
IIT Roorkee has made history by creating MoScNet, an AI model that converts handwritten Modi script—used extensively in medieval Maharashtra—into Devanagari. This advancement marks the first time AI has been successfully trained to handle the entire script rather than isolated characters.
The core of the project is MoDeTrans, a unique dataset of 2,043 real-world Modi manuscript images, annotated with expert-verified Devanagari text. These samples span three historical periods—Shivakalin, Peshwekalin, and Anglakalin—allowing the model to learn from varied handwriting styles.
MoScNet uses a Vision–Language Model (VLM) with knowledge distillation, making it both accurate and lightweight—ideal for deployment in low-resource settings. Tests show it significantly outperforms traditional OCR systems in converting entire lines of text, not just individual characters.
This innovation is especially important because more than 40 million documents—including land records, Ayurvedic texts, and medieval science materials—still exist in Modi script, many deteriorating and inaccessible. With only a few specialists able to read Modi, MoScNet opens up vast potential for historians, researchers, and archivists worldwide.
The model and dataset are fully open-source on Hugging Face, encouraging integration with national missions like Digital India, Bhashini, BharatGPT, and the National Language Translation Mission. IIT Roorkee emphasizes this as a vital step in preserving India’s linguistic and cultural heritage, democratizing access to centuries-old wisdom through AI.
In summary, MoScNet revolutionizes heritage digitization by automating the difficult task of transliterating ancient Modi script into modern Devanagari. This tech leap preserves invaluable manuscripts, supports academic research, and furthers India’s digital language initiatives—while laying groundwork for applying similar AI solutions to other endangered scripts globally.
