Vāgdhenu (वाग्धेनु) — In-Browser Sanskrit Śloka-to-Chant ONNX Weights (sanskrit-tts-onnx)
🌐 Live Web App: https://h3manth.com/ai/sanskrit-tts/ 🤗 Interactive Hugging Face Space: https://huggingface.co/spaces/gnumanth/sanskrit-tts-web 💻 GitHub Pages Demo: https://hemanth.github.io/sanskrit-tts-web/ 📦 Source Code: https://github.com/hemanth/sanskrit-tts-web
Quantized ONNX Runtime Web (WebGPU + WASM SIMD) model weights for Vāgdhenu, a 100% client-side, vṛtta-aware Sanskrit śloka-to-chant neural synthesis engine.
Model Artifacts (~237 MB Total)
| File | Size | Description |
|---|---|---|
vagdhenu_cond_q8.onnx |
20.6 MB | Static Conditioner — 4-layer ConvNeXt V2 text encoder + Rotary Position Embeddings (RoPE) + Rank-32 SVD-factored static bias projection (1039 tokens → 1024-d). Runs once per hemistich. |
vagdhenu_dit_step_q8.onnx |
191.6 MB | 22-Block Flow-Matching DiT Step — Dynamic uint8 weight-quantized Diffusion Transformer with single-node FastGelu, Rank-32 SVD AdaLN modulation, and 0.9999 cosine similarity to FP32. |
vagdhenu_vocos_q8.onnx |
55.6 MB | Fourier-Hann Vocos Vocoder — ConvNeXt backbone (q8) + exact FP32 log-magnitude/phase head + 1026-bin Fourier-Hann ConvTranspose1d overlap-add synthesis (24 kHz). |
baked_bank.bin + baked_bank.json |
2.84 MB | Pre-Baked Prosody Reference Bank — IEEE-754 FP16 mel templates across 18 classical Sanskrit vṛttas (Anuṣṭubh, Śārdūlavikrīḍita, Vasantatilakā, Mandākrāntā, Bhujaṅgaprayāta, etc.), 3 alliteration primes, and the 1,039-token phoneme vocabulary. |
Credits
- 100% In-Browser WebGPU & WASM Architecture: Hemanth HM — h3manth.com/ai/sanskrit-tts
- Original Vāgdhenu Research & Training: Prof. Prathosh A P, IISc Bengaluru