Vāgdhenu (वाग्धेनु) — Fast WASM SIMD (MatMulInteger QUInt8) Sanskrit TTS Weights (sanskrit-tts-wasm)
🌐 Live Web App: https://h3manth.com/ai/sanskrit-tts/ 🤗 Interactive Hugging Face Space: https://huggingface.co/spaces/gnumanth/sanskrit-tts-web 🖥️ WebGPU / ONNX Weights (
sanskrit-tts-onnx): https://huggingface.co/gnumanth/sanskrit-tts-onnx 📦 Source Code & npm Module: https://github.com/hemanth/sanskrit-tts-web
Native WASM SIMD MatMulInteger (QUInt8) quantized weights for Vāgdhenu, a 100% client-side, vṛtta-aware Sanskrit śloka-to-chant neural synthesis engine.
Why sanskrit-tts-wasm is the Default Browser Backend
- 3.24× Faster ODE Step Latency (
257 msvs835 msper step on CPU/WASM SIMD via hardwarei8x16unsigned integer dot-product kernels). - 7.6× Faster Session Initialization (
271 msvs2,066 ms) with 9× lower peak RAM (~280 MBvs2.6 GB), eliminating CPU ConstantFolding weight expansion on mobile and desktop browsers.
Model Artifacts
| File | Size | Description |
|---|---|---|
vagdhenu_cond_q8.onnx |
19.8 MB | WASM INT8 Static Conditioner — QUInt8 MatMulInteger 4-layer ConvNeXt V2 text encoder + RoPE + Rank-32 static bias projection. |
vagdhenu_dit_step_q8.onnx |
199.1 MB | WASM INT8 22-Block Flow-Matching DiT Step — DynamicQuantizeLinear + MatMulInteger (QUInt8) Diffusion Transformer with FastGelu and Rank-32 SVD AdaLN. |
vagdhenu_vocos_q8.onnx |
55.6 MB | Fourier-Hann Vocos Vocoder — ConvNeXt backbone + exact FP32 log-magnitude/phase head + 1026-bin Fourier-Hann ConvTranspose1d 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, 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