Vāgdhenu: Meter-Aware TTS That Chants Sanskrit at Expert-Level Fidelity

Prof. Prathosh A P of the Indian Institute of Science, Bengaluru has built Vāgdhenu, a text-to-speech system that renders Sanskrit verse as traditional pārāyaṇa chant. Paste a śloka in any Indian script and the system automatically detects its meter (vṛtta) and selects a matched reference recitation. The engine is a flow-matching TTS backbone retrained on a purpose-recorded, single-speaker Sanskrit chant corpus of roughly five hours, followed by a voice-steering pass, with the neural vocoder separately fine-tuned for the chant register.

The interesting engineering sits in the linguistics rather than the model size. A script-aware frontend routes all Sanskrit through Kannada orthography to sidestep the schwa-deletion that Devanagari triggers in Hindi-trained pipelines, and the phonology is handled with unusual care: visarga sandhi with its jihvāmūlīya and upadhmānīya allophones, the aspiration contrast, three distinct sibilants and the full retroflex series, homorganic anusvāra, and vocalic ṝ. The result reaches an expert mean-opinion score around 4.6 and correctly pronounces dense conjuncts, including retroflex aspirates that earlier architectures could not handle.

The system is more than a demo. It generated two large chanted corpora — the 5,183-verse Mahābhārata Tātparya Nirṇaya (~17.5 hours) and the roughly 18,000-verse Śrīmad Bhāgavatam — and powers two free apps: Bhāgavatavāṇī, an offline, ad-free Bhāgavatam reader with synced audio and line-by-line karaoke highlighting across ten scripts, and Vāgbodhinī, a chant tutor that scores a learner’s recitation syllable by syllable against Vāgdhenu’s reference. It’s a notable case of speech AI applied to preserving and teaching a classical liturgical tradition.