Vāgdhenu: A Sanskrit Chanting TTS System
a day ago
- #Sanskrit TTS
- #chant tutor
- #meter detection
- Developed a vṛtta (meter) aware śloka-to-chant text-to-speech system for Sanskrit, allowing users to paste verses in any Indian script for automatic meter detection.
- First chant takes 10–60 seconds for model warm-up; offers a backup demo and renders six vṛttas, including verses from shipped deployments.
- System powers a free app for the complete Śrīmad Bhāgavatam with synced audio, karaoke highlighting, offline functionality, and search in 10 Indian scripts.
- Companion tool, Vāgdhenu, serves as a Sanskrit chant tutor, providing reference chants and scoring syllables to help users improve their chanting.
- Vāgdhenu uses a flow-matching TTS backbone retrained on a single-speaker Sanskrit chant corpus (~5 hours), with voice-steering and fine-tuned neural vocoder.
- Includes script-aware frontend through Kannada orthography, handling visarga sandhi, aspiration contrast, distinct sibilants, retroflex series, and meter detection under the half-reference rule.
- Achieves expert MOS of about 4.6, correctly rendering dense conjuncts and retroflex aspirates that earlier architectures failed to handle.
- Generated large-scale corpora: Mahābhārata Tātparya Nirṇaya (32 chapters, 5,183 verses) and Śrīmad Bhāgavatam (~18,000 verses across 12 books).