Vāgdhenu — Sanskrit Chant TTS

"The wish-cow of speech." A production-grade, single-speaker Sanskrit chant (pārāyaṇa) text-to-speech system — it chants classical ślokas with metrically-aware durations and tradition-faithful melodic contour, not flat read-aloud.

MOS ~4.6 (expert listener). Conjuncts — including retroflex aspirates (ṣṭ, ḍḍh, …) — render 100% correctly, the class earlier architectures could not crack. Used to produce MBTN (32 YouTube videos, 17h 34m) and the Śrīmad Bhāgavatam (16,017 verses, audio app + 31 karaoke videos).

[ Project page + live demo · Model weights → HF · Demo → HF Space · Tech report → docs/TECH_REPORT.md ]

Demos (rendered with this system)

• Mahābhārata Tātparya Nirṇaya (MBTN) — full chant series: YouTube playlist
• Śrīmad Bhāgavatam — karaoke-video series: YouTube playlist

Developed and maintained by Prof. Prathosh, Indian Institute of Science, Bengaluru.

How it works

• Backbone: IndicF5 / F5-TTS — a flow-matching DiT (OT-CFM mel-infilling, ~337M params, no native duration or pitch head). Sanskrit is routed through Kannada script (Devanagari triggers Hindi schwa-deletion).
• Vocoder: NVIDIA BigVGAN-v2, fine-tuned on F5 vocos-mel (mandatory — vocos shivers on long vowels).
• Prosody: F5's content fidelity is bulletproof but its prosody is text-driven, not designable. The working levers are the reference clip (voice + swara + pace, via the half-reference rule) and a voice-steering fine-tune. (See docs/TECH_REPORT.md §14 for the full account — this is the central architectural finding.)
• Text frontend (src/prep_text.py) — the most reusable piece: Deva→SLP1→Kannada routing, internal visarga sandhi (utva/rutva/lopa/satva), homorganic anusvāra, vocalic-ṝ handling, daṇḍa-final rules, meter/gaṇa (L/G) detection.

Layout

src/         text frontend, meter detection, inference, post-gate, reference bank
pipeline/    data-prep (cut→pair→train) + build/assemble/QC
demo/        Gradio app (HF ZeroGPU)
docs/        scrubbed technical report + frontend/pipeline references
examples/    sample inputs + rendered outputs
scripts/     env setup + weight download

Install & quickstart

Requires Python 3.10 and a CUDA 12.1 GPU.

bash scripts/setup.sh    # torch+cu121, deps, BigVGAN, and downloads weights -> models/
# render a Devanagari verse (+ meter) to a chanted wav:
python src/render.py --shard examples/sample_shard.json --results /tmp/res.json --outdir out
# -> out/sample_anushtubh.wav

The batch renderer takes a shard JSON: [{"id","meter","padas":[devanagari…],"seed","out"}]. For one-off single-verse renders see src/render_production.py. CHAMP_ROOT env overrides the weights dir (default models/).

Case studies

• MBTN (Mahābhārata Tātparya Nirṇaya) — 32-adhyāya video deliverable (Devanagari + Kannada karaoke, tanpura), shipped.
• Śrīmad Bhāgavatam — 12 skandhas, ~18k verses, audio app + a 31-video 3-script (Devanāgarī · Kannada · IAST) karaoke series. Sanskrit text gratefully acknowledged to Poornaprajna Samshodhana Mandiram, Bengaluru.

Attribution & licenses

• Code: Apache-2.0 (LICENSE).
• Built on AI4Bharat IndicF5 (MIT), NVIDIA BigVGAN-v2, and F5-TTS — see their licenses; weights redistributed per those terms.
• Model weights + intended-use/ethics note: see the HF model card.

Ethics / intended use

Single-speaker synthesis of sacred Sanskrit recitation, for pārāyaṇa/study/accessibility. The voice is the author's own. Please use responsibly; do not impersonate.

Citation

(BibTeX added with the arXiv report.)