Vāgdhenu — "the wish-cow of speech" — is a single-speaker Sanskrit chant (pārāyaṇa) text-to-speech system that has drawn 465 stars and 109 forks since its 28 June 2026 release, with the weights pulling 15,025 downloads on Hugging Face. What separates it from the crowded TTS field is the target: it does not read ślokas aloud, it chants them, with metrically-aware durations and a tradition-faithful melodic contour. It is developed and maintained by Prof. Prathosh at the Indian Institute of Science, Bengaluru, and the reported quality is MOS ~4.6 from an expert listener.

The most technically interesting claim concerns conjuncts. Sanskrit consonant clusters — including retroflex aspirates such as ṣṭ and ḍḍh — are described as rendering 100% correctly, and the README is explicit that this is the class of sound earlier architectures could not crack. The route to that result is unusual: Sanskrit text is deliberately routed through Kannada script rather than Devanagari, because Devanagari input triggers Hindi schwa-deletion in the underlying model. That kind of orthographic workaround is exactly the sort of practical finding that rarely survives into a published paper but decides whether a system actually works.

The stack builds on existing open components rather than training from zero. The backbone is IndicF5 / F5-TTS, a flow-matching DiT at roughly 337M parameters using OT-CFM mel-infilling — notably with no native duration or pitch head, which is precisely the problem a chant system has to solve. The vocoder is NVIDIA BigVGAN-v2, fine-tuned on F5 vocos-mel and described as mandatory because vocos "shivers" on the long vowels that Sanskrit recitation is full of. On prosody, the project reports a candid architectural finding: F5's content fidelity is bulletproof but its prosody is text-driven, not designable, so the only working levers are the reference clip (carrying voice, swara, and pace via a "half-reference rule") and a voice-steering fine-tune. The technical report treats this as the central result.

The piece most likely to be reused elsewhere is the text frontend (src/prep_text.py), which handles Devanāgarī→SLP1→Kannada routing, internal visarga sandhi (utva, rutva, lopa, satva), homorganic anusvāra, vocalic-ṝ handling, daṇḍa-final rules, and meter/gaṇa (L/G) detection. That is a substantial piece of Sanskrit computational linguistics shipped as working code, useful well beyond this one model.

The project is validated by shipped output rather than benchmarks alone. It produced MBTN (Mahābhārata Tātparya Nirṇaya) as 32 YouTube videos totalling 17 hours 34 minutes, and the Śrīmad Bhāgavatam at 16,017 verses across 12 skandhas as an audio app plus a 31-video three-script karaoke series. Rendering is batch-oriented through a shard JSON, with src/render_production.py for one-off verses and a Gradio demo on HF ZeroGPU.

The constraints are narrow by design. It is single-speaker only — the voice is the author's own — and the ethics note asks that it be used for pārāyaṇa, study, and accessibility, not impersonation. It requires Python 3.10 and a CUDA 12.1 GPU, so there is no CPU or Apple-silicon path. Licensing is layered: the code is Apache-2.0, but it inherits from AI4Bharat IndicF5 (MIT), NVIDIA BigVGAN-v2, and F5-TTS, with weights redistributed under those terms. Last pushed 19 July 2026, and the citation BibTeX is still pending an arXiv report.