Vāgdhenu: Sanskrit Chanting TTS System Launched
Original: Vāgdhenu: A Sanskrit Chanting TTS System
Why This Matters
Demonstrates production-scale TTS for classical Sanskrit, a low-resource language with complex phonological requirements.
Prof. Prathosh A P of Indian Institute of Science, Bengaluru has released Vāgdhenu, a meter-aware text-to-speech system for Sanskrit chanting. Trained on ~5 hours of single-speaker chant audio, it achieves an expert MOS of 4.6 and has rendered over 23,000 verses across two major Sanskrit corpora.
Vāgdhenu is a vṛtta (meter) aware śloka-to-chant TTS system for Sanskrit developed by Prof. Prathosh A P at the Indian Institute of Science, Bengaluru. The system uses a flow-matching TTS backbone retrained on a purpose-recorded single-speaker Sanskrit chant corpus of approximately 5 hours, with a fine-tuned neural vocoder optimized for the chant register. It achieves an expert Mean Opinion Score (MOS) of approximately 4.6.
The pipeline includes a script-aware frontend routing Sanskrit through Kannada orthography to avoid Hindi schwa-deletion artifacts from Devanagari, and correctly handles visarga sandhi, aspiration contrasts, three sibilants, the full retroflex series, homorganic anusvāra, and vocalic ṝ. A meter-detection mechanism selects a matched reference chant pattern using the half-reference rule.
The system has been deployed at scale, generating approximately 17.5 hours of audio for the Mahābhārata Tātparya Nirṇaya (5,183 verses) and chanting for approximately 18,000 verses of the Śrīmad Bhāgavatam. The latter powers Bhāgavata-VāNi, a free offline app available on Android and iOS supporting 10 Indian scripts with karaoke-style line highlighting. A companion tool, Vāgbodhinī, uses the system as a reference for Sanskrit pronunciation coaching. Code is released under Apache-2.0, data under CC-BY-4.0, using AI4Bharat IndicF5 and NVIDIA BigVGAN-v2 base models.