Nandini: A Female Voice for Vagdhenu, a Vrtta (Meter) Aware Sloka-to-Chant (TTS) System for Sanskrit
Year: 2026
Type: preprint
Abstract: We present Nandini, a female chanting voice for Vagdhenu. Vagdhenu turns a Sanskrit verse into chant that follows its meter, in the voice of a single male reciter. A female voice would normally need recordings of a trained female reciter, but Nandini is built only from Vagdhenu’s released data. Each of its 1,467 recorded verses is converted into a voice taken from a single 25-second text-to-speech sample, and the reciter’s melody is then put back in the new voice’s range. Up to ten such candidates are made for each verse, and the first in which a chant recogniser finds at most one syllable wrong is kept. By that recogniser’s count, 1,298 of the 1,467 clips (88.5%) meet this bar, which is optimistic because the same recogniser chose them. IndicF5 is then fine-tuned on the clips with Vagdhenu's recipe, together with a vocoder for the new voice. On 60 verses new to both systems, Nandini's character error rate is 3.81% against Vagdhenu’s 3.67%, and its chant is 4.5% longer. By ear, its tune matches Vagdhenu’s fully or mostly in 5 of 8 blind pairs. In the other 3 the melody is also similar, and only the pauses fall in different places. We release the dataset, the code and the model, including an ONNX version
Author: Athreya, Ram G
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7237350452
Cites:
Cited by:
Related to:
