Introduction

Ninada just leveled up for makers who want Indic and Classical Sanskrit speech in the browser. On 2026-11-05, Hemanth HM cut v5.1.0 — dual-tier 24 kHz streaming specialists — on gnumanth/ninada (lastModified ~09:15 UTC) with the matching GitHub release commit the same morning, and a same-day live studio at hemanth.github.io/ninada.

This is generative text-to-speech, not an ASR fine-tune: dedicated male/female voices per language, routed automatically between desktop HiFi-GAN and phone iSTFTNet weights.

What Shipped

  • Vendor / product: Hemanth HM — Ninada (निनाद) v5.1.0
  • Modality: open-weight in-browser / Node / Python text-to-speech for Classical Sanskrit (sa), English (en), and 10 Indic languages (hi, kn, mr, bn, ta, te, ml, gu, pa, or)
  • Open vs closed: runtime MIT; model weights CC-BY-4.0 (attribute AI4Bharat Rasa / Hi-Fi TTS / IISc SYSPIN priors as listed on the card)
  • Runnable today: open hemanth.github.io/ninada (no signup), or npm i ninada / pip install ninada against the HF specialists
  • Concrete fact: 48 ONNX checkpoints — 24 studio ~27.9 MB HiFi-GAN + 24 mobile ~21.7 MB iSTFTNet; on a Google Pixel 10 Pro XL (Chrome WASM) the card reports 519 ms first-chunk compute and ~0.66× steady vocoder RTF with 0 AudioWorklet underruns

v5.1.0’s headline is capability routing: desktops keep the full studio HiFi-GAN path; phones auto-load the distilled mobile vocoder instead of fighting WebGPU cold starts.

Run It

Browser smoke test: open hemanth.github.io/ninada and pick a language + male/female voice.

JS streaming sketch:

import ninada from 'ninada';

const voice = await ninada('gnumanth/ninada', {
  lang: 'sa',
  voice: 'female',
});

for await (const chunk of voice.stream('नमस्ते')) {
  playPcm24k(chunk.pcm, chunk.sampleRate); // 24 kHz float32
}

Python path: from ninada import Ninada then Ninada(lang="kn", voice="male").stream(...). Weights live under onnx/ (studio) and onnx_mobile/ (phone) on the HF card.

Notes for makers

  • Each language is a separate specialist — the card’s pitch is zero cross-language accent bleed versus one giant multilingual checkpoint
  • Studio English uses clean Hi-Fi TTS speakers; Indic languages train on AI4Bharat Rasa 24 kHz studio recordings
  • Earlier Ninada drops (Oct 31) were the initial fleet; v5.1.0 is the dual-tier mobile unlock worth covering today
  • Distinct from ArtRealmAI’s older Indic Tiny TTS piece — different vendor, browser dual-tier vocoders, Sanskrit + 11 other locales
  • No Gen CTA: browser Indic TTS does not map to the Flux / Grok / MiniMax-H3 / Seedream / Seedance Gen loop

Conclusion

If you need Sanskrit or Indic speech that stays on-device in a tab — studio quality on desktop, phone-friendly iSTFTNet on mobile — Ninada v5.1.0 is today’s runnable drop: 48 ONNX specialists, live studio, paste-ready npm/pip.

—Anabel ♡