Ninada open-sources 48 in-browser 24kHz TTS models for Sanskrit and 10 Indic languages, 538ms on phone WASM

init0 · reddit · 2026-11-05

Ninada is open-sourced: 48 in-browser 24 kHz TTS specialist models (MIT code, CC-BY-4.0 ONNX checkpoints) covering Classical Sanskrit, English and 10 Indic languages, running 100% locally with no sign-up and 538ms first audio on phone WASM.

Two counterintuitive findings from real-phone profiling (Pixel 10 Pro XL):

Mobile WebGPU lost to WASM: Dawn shader compilation stalled cold start 6-14s, and GPU buffer readback pushed synthesis slower than real time (1.48x RTF)
2-3 WASM threads ran slower than one: big.LITTLE scheduling stalled ONNX Runtime spin barriers (1.76x RTF at 2 threads vs 1.19x at 1), recovering only at 4 threads

Key engineering:

48 native-script specialists with mixed W8A32 encoder/flow + FP16 vocoder quantization, 0.00 dB MCD penalty
Device-routed dual-tier vocoders: desktops get a 27.9MB HiFi-GAN (89ms WebGPU audible start), phones auto-route to a 21.7MB distilled iSTFTNet with 2.64x fewer vocoder MACs
A <15ms in-browser ONNX protobuf splitter enables streaming, cutting browser RAM 72% (584MB → 162MB)
3.09-3.22 dB held-out MCD (N=480, bootstrap 95% CIs), offline CacheStorage persistence

Live demo, GitHub and Hugging Face weights are all available.

Original post →

Ninada open-sources 48 in-browser 24kHz TTS models for Sanskrit and 10 Indic languages, 538ms on phone WASM

More from Multimodal