{"slug": "vagdhenu-open-source-meter-aware-text-to-speech-for-sanskrit", "title": "Vāgdhenu: Open-source meter-aware text-to-speech for Sanskrit", "summary": "Vāgdhenu, an open-source meter-aware text-to-speech system for Sanskrit, automatically detects verse meters and generates chant recitations. Developed by retraining a flow-matching TTS backbone on a custom 5-hour Sanskrit chant corpus, it achieves an expert MOS of 4.6 and correctly renders complex conjuncts. The system has produced two large-scale corpora: the Mahābhārata Tātparya Nirṇaya (5,183 verses) and the Śrīmad Bhāgavatam (~18,000 verses).", "body_md": "Paste a Sanskrit verse in any Indian script — the meter is detected automatically.\n\nFirst chant takes ~10–60s while the model warms up. If the demo doesn't load,\nuse the [backup demo ↗](https://prathoshap-vagdhenu-demo.hf.space?__theme=light).\n\nSix vṛttas rendered by this system — including verses from the shipped deployments.\n\nVāgdhenu maps a metrical verse to its chanted *pārāyaṇa* recitation. Its voice is a\nflow-matching TTS backbone **retrained on a purpose-recorded, carefully designed\nsingle-speaker Sanskrit chant corpus** (~5 hours), with a further voice-steering retrain;\nthe neural vocoder is likewise fine-tuned for the chant register. Around the trained model sits\nthe machinery a faithful Sanskrit chant pipeline needs: a script-aware frontend that routes\nSanskrit through Kannada orthography (avoiding the Hindi schwa-deletion that Devanagari triggers);\nvisarga sandhi with the jihvāmūlīya and upadhmānīya allophones; the aspiration contrast; the three\nsibilants and the full retroflex series kept distinct; homorganic anusvāra and vocalic ṝ; and a\nvṛtta-aware mechanism that detects the meter and selects a matched reference under the\n*half-reference rule*. The retrained model reaches an expert MOS of about **4.6**,\nand dense conjuncts — including retroflex aspirates — render correctly, the class earlier\narchitectures could not crack.\n\nThis system produced two corpora at scale.\n\n● **Mahābhārata Tātparya Nirṇaya** — 32 chapters, 5,183 verses (~17.5h) ·\n[video series ↗](https://www.youtube.com/playlist?list=PLL1s8qiaGy0IP0G_PhlwaGA5EOfzoKrV_)\n\n● **Śrīmad Bhāgavatam** — ~18,000 verses across 12 books ·\n[karaoke-video series ↗](https://www.youtube.com/playlist?list=PLDiYyVdyo2Sc)", "url": "https://wpnews.pro/news/vagdhenu-open-source-meter-aware-text-to-speech-for-sanskrit", "canonical_source": "https://prathosh.in/vagdhenu/", "published_at": "2026-06-30 05:33:55+00:00", "updated_at": "2026-06-30 05:50:13.307944+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "generative-ai", "ai-research", "ai-products"], "entities": ["Vāgdhenu", "Mahābhārata Tātparya Nirṇaya", "Śrīmad Bhāgavatam"], "alternates": {"html": "https://wpnews.pro/news/vagdhenu-open-source-meter-aware-text-to-speech-for-sanskrit", "markdown": "https://wpnews.pro/news/vagdhenu-open-source-meter-aware-text-to-speech-for-sanskrit.md", "text": "https://wpnews.pro/news/vagdhenu-open-source-meter-aware-text-to-speech-for-sanskrit.txt", "jsonld": "https://wpnews.pro/news/vagdhenu-open-source-meter-aware-text-to-speech-for-sanskrit.jsonld"}}