prathoshap/sushrota-sanskrit-asr

Su-shrota is a research project that builds speech recognition for Sanskrit, specifically tuned for the formal, chanted style used in religious and scholarly recitation rather than everyday conversation. It also includes a companion tool called Vagbodhini, an interactive practice app that helps people improve their chanting by checking how closely their spoken recitation matches a known reference text. The project is built around a machine learning model that was originally trained on general Sanskrit speech and then further trained, or fine tuned, specifically on recordings of chants and formal prose. The README documents not only the version that shipped, but also many experiments that did not work, including attempts to train on the model's own guesses, various advanced training tricks, and automatic text correction tools, all of which turned out to make results worse in careful testing. One of the project's central findings is that word boundaries in written Sanskrit are often just a spelling convention, since spoken Sanskrit blends words together in ways that written text does not always separate cleanly. Because of this, the researchers created their own way of scoring accuracy that ignores those spacing differences, since a large share of what looks like an error is really just a word-boundary mismatch rather than an actual mistake in what the model heard. The most successful approach turned out to be collecting real recordings from people using the public chanting and dictation tools, with their consent, and using that everyday audio to keep improving the model over time. This closed a large part of the gap between how well the model performed on clean studio recordings versus messy real-world audio, without making the studio results worse. The tools are live on the web for anyone to try, and the underlying code and trained model weights are shared publicly. This project is aimed at Sanskrit scholars, students of Sanskrit recitation, and researchers interested in speech recognition for low-resource or classical languages.