Vāgdhenu — Sanskrit Chant TTS (weights)
Model weights for Vāgdhenu, a single-speaker Sanskrit chant (pārāyaṇa) TTS. MOS ~4.6 (expert listener); conjuncts including retroflex aspirates render 100% correctly.
Files
| file | what |
|---|---|
voice_steer_ema_2026-06-17.pt |
Production voice — voice-steered, more reference-responsive (the recommended default). |
voice_armA_ema_2026-06-11.pt |
Fallback — reference-driven chant (voice + swara + pace from the reference clip). |
voc_bigvgan_EMA_2026-06-11.pth |
Vocoder — NVIDIA BigVGAN-v2 fine-tuned on F5 vocos-mel (mandatory; vocos shivers on long vowels). |
The base DiT + vocab.txt come from ai4bharat/IndicF5 (auto-downloaded by the repo's setup).
Architecture
IndicF5 / F5-TTS — flow-matching DiT (OT-CFM mel-infilling, dim 1024 / depth 22 / heads 16, ~337M params, no native duration or pitch head) → BigVGAN-v2 vocoder. Sanskrit is routed through Kannada script (Devanagari triggers Hindi schwa-deletion).
Usage
See the GitHub repo — bash scripts/setup.sh downloads these weights to models/, then python src/render.py ... renders a Devanagari verse + meter to a chanted wav.
The key lever is the reference (F5 prosody is text-driven, not designable): supply the prosody you want as a clean, exactly-matched reference clip (the half-reference rule — ref_text must match the reference audio's spoken span on a word/daṇḍa boundary). A per-meter reference bank ships with the repo.
Training
Fine-tuned from IndicF5 on a ~5 h single-speaker Sanskrit chant corpus (prathoshap/vagdhenu-data); the production voice adds a voice-steering retrain on paired clips. Full method in the repo's docs/TECH_REPORT.md.
Intended use & limitations
Synthesis of classical Sanskrit chant (pārāyaṇa) for recitation, study, and accessibility. No Vedic svaras. Prosody is reference-driven, not arbitrarily designable. The voice is the author's own — please use responsibly and do not impersonate.
License & attribution
Our contribution under Apache-2.0. Built on AI4Bharat IndicF5 (MIT), NVIDIA BigVGAN-v2, and F5-TTS — the vocoder is a BigVGAN-v2 derivative; please observe NVIDIA's BigVGAN license terms. Cite Vāgdhenu (BibTeX with the technical report).