ninada

In-browser, Node.js, and Python 24,000 Hz text to speech for English, Classical Sanskrit, and ten Indian languages (sa, en, hi, kn, mr, bn, ta, te, ml, gu, pa, or) powered by capability-routed dual-tier 24 Male & Female studio specialists (27.9 MB HiFi-GAN v1 on Desktop/WebGPU and 21.7 MB c8c8_384 iSTFTNet on Mobile WASM) with zero cross-language accent bleed.

npm install ninada webml-kit

pip install ninada

Quick start

import ninada from 'ninada';
import { detectDevice } from 'webml-kit';

const { backend } = await detectDevice();
const voice = await ninada('gnumanth/ninada', {
  lang: 'kn',
  voice: 'male',
  device: backend,
});

for await (const chunk of voice.stream('ನಮಸ್ಕಾರ, ಇಂದು ಬೆಂಗಳೂರಿನಲ್ಲಿ ಹವಾಮಾನ ತುಂಬಾ ಚೆನ್ನಾಗಿದೆ.')) {
  playPcm24k(chunk.pcm, chunk.sampleRate); // Float32Array @ 24,000 Hz
}

const wav = await voice.speak('ಕನ್ನಡವೇ ಸತ್ಯ ಕನ್ನಡವೇ ನಿತ್ಯ', { lang: 'kn', voice: 'female' });
await voice.dispose();

voice.stream() yields 24,000 Hz PCM clauses as they synthesize with 0 AudioWorklet underruns and CacheStorage + IndexedDB persistence (0 refetches on reload). voice.speak() returns a complete 24 kHz RIFF WAV buffer.

Python

from ninada import Ninada

tts = Ninada(lang="sa")
for chunk in tts.stream("प्रज्ञापूर्वकव्यापारवत्त्वं जीवत्वम् । यथार्थज्ञानं प्रमाणम् ॥"):
    play(chunk["pcm"], chunk["sample_rate"])  # float32 @ 24,000 Hz (~0.197x CPU RTF)

wav_bytes = tts.speak("सत्यमेव जयते नानृतम्", lang="sa", voice="female")
open("satyam.wav", "wb").write(wav_bytes)

Ninada automatically routes by device capability between 27.9 MB Studio HiFi-GAN v1 ONNX specialists (onnx/{sa,en,hi,kn,mr,bn,ta,te,ml,gu,pa,or}.onnx and onnx/*_female.onnx) for Desktop/WebGPU and 21.7 MB Mobile c8c8_384 iSTFTNet ONNX specialists (onnx_mobile/*.onnx, 0.66x mobile WASM RTF) for phones and low-memory CPUs, each trained exclusively on its own language's ai4bharat/Rasa (CC-BY-4.0) or MikhailT/hifi-tts (CC-BY-4.0) single-speaker 24 kHz studio recordings and native Unicode orthography.

CLI

npx ninada "Welcome to Ninada studio voice synthesis." -l en -v female -o english.wav
ninada "सत्यमेव जयते नानृतम्" -l sa -o satyam.wav

Runs 24 kHz neural synthesis when ONNX weights are available and falls back automatically to a 0.0 MB zero-dependency Cascade Klatt-Akṣara synthesizer offline.

Native Script Orthography & Prosody

import { studioTextToIds, textToIds } from 'ninada';

const { studioIds, lang } = studioTextToIds('मराठी भाषेची साहित्य परंपरा अत्यंत समृद्ध आहे.', 'mr');
const { aksharas } = textToIds('धर्मक्षेत्रे कुरुक्षेत्रे समवेता युयुत्सवः ।', 'sa');

studioTextToIds() tokenizes directly in each language's native Unicode character table (sa 144 tokens, en 102, hi 120, kn 85, mr 144, bn 100, ta 61, te 110, ml 101, gu 124, pa 85, or 94), preserving Classical Sanskrit vocalic ṛ/ṝ/ḷ, visarga sandhi, English studio prosody, Kannada ಙ/ಃ allophony, Tamil ழ/ற/ன, Malayalam atomic chillus (ൺ ൻ ർ ൽ ൾ ൿ), Gurmukhi tippi/addak (ੰ/ੱ), and Odia/Bengali conjuncts feeding a 24 kHz RelativePositionTransformer + StochasticDurationPredictor paired with HiFi-GAN v1 (Studio) and c8c8_384 iSTFTNet (Mobile).

Benchmarks (N = 480 Held-Out 24 kHz Studio Utterances, B = 1,000 Bootstrap 95% CIs)

System	Rate	Footprint (↓)	Coverage (↑)	TTFA p50 (↓)	MCD 0–8 kHz (dB, ↓)	8–12 kHz HF Err (dB, ↓)	Dur MAE (↓)
Base facebook/mms-tts-hin	16 kHz	138.8 MB	29.4%	0.9 ms*	7.98 [7.81, 8.16]	65.16 dB	3.49s
Meta MMS 11-Specialist Fleet (11x VITS)	16 kHz	1526.8 MB	100.0%	204.2 ms	5.17 [5.12, 5.22]	64.97 dB	1.04s
ninada v5.1.0 Untuned Warm-Start (step=0)	24 kHz	112.0 MB	100.0%	191.0 ms	4.58 [4.54, 4.62]	10.93 dB	0.93s
ninada v5.1.0 Native Studio FP32	24 kHz	112.0 MB	100.0%	43.3 ms	3.22 [3.19, 3.25]	6.08 dB	0.58s
ninada v5.1.0 Studio ONNX (HiFi-GAN)	24 kHz	27.9 MB	100.0%	43.3 ms	3.22 [3.19, 3.25]	6.07 dB	0.58s
ninada v5.1.0 Mobile ONNX (iSTFTNet)	24 kHz	21.7 MB	100.0%	519.0 ms	3.34 (1.34 dB distill)	6.28 dB	0.58s
ninada Tier-2 Klatt Fallback	24 kHz	0.0 MB	100.0%	110.2 ms	10.53 [10.43, 10.64]	42.38 dB	0.88s

Per-language held-out MCD 0–8 kHz (dB, ↓): Punjabi 2.89 (vs 5.23 MMS), Indian English 3.07 (vs 5.94 MMS), Marathi 3.10 (vs 4.62 MMS), Odia 3.11 (vs 5.49 MMS), Tamil 3.21 (vs 4.87 MMS), Bengali 3.27 (vs 4.68 MMS), Gujarati 3.30 (vs 5.56 MMS), Malayalam 3.35 (vs 4.88 MMS), Hindi 3.36 (vs 4.85 MMS), Kannada 3.36 (vs 5.04 MMS), Sanskrit 3.51 (vs 5.32 MMS), Telugu 3.59 (vs 5.57 MMS). Full bootstrap CIs and parity telemetry (0.58 dB parity MCD, r = 0.9999) in bench/SCIENCE_REPORT.md and bench/results.json.

Reproduce & Verify

PYTHONPATH=python/src:. python training/train_native_specialists_gcp.py --work-dir data/v5 --out-dir artifacts/ninada_v5
PYTHONPATH=python/src:. python bench/check_bench_thresholds.py
pytest -q && node verify-browser.mjs

Trains all 24 native 24 kHz Male/Female studio VITS specialists across 12 languages, evaluates 480 held-out studio clips, and verifies 0 AudioWorklet underruns and 0 cached network refetches in headless Chromium.

Links

• Paper (arXiv:2611.10035)
• In-Browser Voice Studio
• Hugging Face Model (gnumanth/ninada)
• Hugging Face Dataset (gnumanth/ninada-sanskrit-indic)

Citation

@article{hm2026ninada,
  title   = {Ninada: Sub-Clause Split-Graph VITS Streaming and Capability-Routed 24 kHz Speech Synthesis Across Twelve Scripts},
  author  = {HM, Hemanth},
  journal = {arXiv preprint arXiv:2611.10035},
  year    = {2026},
  url     = {https://arxiv.org/abs/2611.10035}
}

License

MIT © Hemanth.HM