And here I am opening sourcing Vaghenu, a meter aware sloka-to-chant, TTS for Sanskrit . This is the world's first vrutta-aware, open-source TTS for Sanskrit Chanting. I am making the model weights, training scripts, and even data (that I meticulously collected) public - https://prathosh.in/vagdhenu/
No large AI lab. No big engineering team. No venture-scale budget. Just a professor's conviction that one of humanity's oldest knowledge traditions deserves modern, open infrastructure.
The name comes from the Upanishadic phrase: "Vācaṃ dhenum upāsīta" - Like the mythical wish-fulfilling cow, Vāgdhenu is intended to make Sanskrit texts more accessible to students, teachers, researchers, and devotees everywhere.
Test out the live demo here and let me know your comments - https://prathosh.in/vagdhenu/
The entire system, from data collection to model building and demos, is built by a single person (your truly) using the powerful harness that we are building at LatentForce.
I have attached a sample audio file generated by the system.
P.S: Posting on behalf of my friend, their aren't on HN.
sieve•3h ago
The documentation of the `transliteration to Kannada` workflow is good. I (and I think a few others) was using this last year to produce story narrations of Chandamama and other Sanskrit stories using the audio models available on Google AI Studio. The schwa deletion on the Hindi side is infuriating with Sanskrit content.
Also, I wonder if this can support prose? Would be helpful for proofreading efforts because you can focus on the page image rather than constantly switch between two panes.