Ninada (निनाद) v2.0 — Streaming Pan-Indic & Classical Sanskrit Neural TTS

Ninada (Nināda — Sanskrit निनाद, deep resonant sound) v2.0 is a sub-second streaming Classical Sanskrit & Pan-Indic neural text-to-speech suite (54.7 MB active split streaming FP16 ONNX per specialist: 27.3 MB Encoder+SDP+Flow + 27.4 MB Halo-Chunked HiFi-GAN, 180.1 ms median TTFA) supporting 6 Indic languages (Sanskrit [sa], Kannada [kn], Hindi [hi], Tamil [ta], Telugu [te], Malayalam [ml]) across 9 Brahmic scripts + ISO-15919/IAST.

• Live Off-Main-Thread Web Worker Studio: https://hemanth.github.io/ninada/
• GitHub Repository & Streaming SDKs (JS + Python): https://github.com/hemanth/ninada
• Curated Split Metadata (1438 Studio Clips): https://huggingface.co/datasets/gnumanth/ninada-sanskrit-indic

GCP NVIDIA L4 GPU Training & Halo-Crossfaded Streaming Architecture

• Hardware & Objective: Trained on Google Cloud (NVIDIA L4 24 GB GPU) with alternating MPD + MRD (LSGAN + multi-layer discriminator feature-matching) + Flow KL + ONNX-safe Stochastic Duration Predictor (SDP) MLE + validation-only end-to-end WiSE-FT across 6 native-script VITS + HiFi-GAN specialists (specialists/{sa,kn,hi,ta,te,ml}).
• Halo-Crossfaded Streaming (123.7 dB Seam SNR): Splits inference into onnx/{lang}/encoder_flow_quantized.onnx (~27.3 MB, native stochastic duration predictor + normalizing flow) and onnx/{lang}/vocoder_quantized.onnx (~27.4 MB, 16-frame receptive-field halo + 256-sample raised-cosine boundary crossfade), persisted in browser CacheStorage (0 network fetches on reload or language switch).
• Anechoic Male Studio Corpus (1438 clips, 127.55 min at 16 kHz; 1150 train / 144 val / 144 test):
  • Sanskrit (388 Male studio clips): ai4bharat/Rasa neutral Sanskrit studio narration + prathoshap/sushrota-sanskrit-asr-data (recitation_corpus + scholar, studio_align_cer <= 5.0%, synthetic tts_aug excluded).
  • Pan-Indic (1,050 Male studio clips — 210 each across hi, kn, ta, te, ml): ai4bharat/Rasa 48 kHz -> 16 kHz anechoic booth recordings.

Held-Out Test Leaderboard (N = 144 Utterances; B = 1,000 Bootstrap 95% CIs)

System	Active Footprint (↓)	Coverage (↑)	TTFA p50 (ms, ↓)	MCD (dB, ↓)	Log-F0 RMSE (c, ↓)	Dur MAE (s, ↓)	Sanskrit CER (%, ↓)	Sanskrit WER (%, ↓)
Studio Human Audio (Topline)	—	100.0%	—	—	—	—	1.78 [1.14, 2.46]%	19.06 [12.81, 25.70]%
Base facebook/mms-tts-hin (Raw)	138.8 MB	43.1%	0.6 ms	6.50 [6.30, 6.71]	961.1	3.10s	19.48 [16.51, 22.60]%	70.30 [60.81, 79.41]%
Base mms-tts-hin + Ninada G2P	138.8 MB	100.0%	463.4 ms	5.49 [5.40, 5.58]	654.8	0.89s	20.27 [17.07, 23.62]%	70.83 [60.99, 80.35]%
Meta MMS 5-Specialist Fleet (5x Models)	694.0 MB	100.0%	507.5 ms	5.09 [5.02, 5.17]	663.7	1.31s	19.48 [16.51, 22.60]%	70.30 [60.81, 79.41]%
Ninada v2.0 Tier-1A (GCP L4 FP32)	138.4 MB	100.0%	150.8 ms	4.70 [4.63, 4.77]	628.0	0.72s	16.65 [13.17, 20.36]%	57.95 [49.77, 66.54]%
Ninada v2.0 Tier-1B (Streaming ONNX)	54.7 MB	100.0%	180.1 ms	4.69 [4.61, 4.76]	626.4	0.73s	16.95 [13.46, 20.28]%	58.79 [49.94, 67.29]%

Streaming Quickstart

import ninada from 'ninada';

const voice = await ninada('kannada', {
  encoderModel: 'https://huggingface.co/gnumanth/ninada/resolve/main/onnx/kn/encoder_flow_quantized.onnx',
  vocoderModel: 'https://huggingface.co/gnumanth/ninada/resolve/main/onnx/kn/vocoder_quantized.onnx',
});
for await (const chunk of voice.stream('ಕನ್ನಡ ಭಾಷೆಯ ನೈಸರ್ಗಿಕ ಧ್ವನಿ ಸಂಶ್ಲೇಷಣೆಗೆ ಸ್ವಾಗತ.')) {
  playPcm16k(chunk.pcm); // Sub-second TTFA!
}

from ninada import Ninada

tts = Ninada(lang="sa")
for chunk in tts.stream("प्रज्ञापूर्वकव्यापारवत्त्वं जीवत्वम् । यथार्थज्ञानं प्रमाणम् ॥"):
    play(chunk["pcm"], chunk["sample_rate"])