Ninada (निनाद) v5.1.2 - Dual-Tier 24 kHz Studio & Mobile Streaming Text-to-Speech Specialists (12 Languages x 2 Voices)

Ninada (Ninada, Sanskrit निनाद, deep resonant sound) v5.1.2 is a true full-band 24,000 Hz In-Browser & Multi-Runtime Neural Text-to-Speech suite featuring capability-based dual-tier weight routing across 12 dedicated Studio Male + 12 dedicated Studio Female specialists (24 voices, 48 ONNX models total) for Classical Sanskrit (sa), Indian English (en), Hindi (hi), Kannada (kn), Marathi (mr), Bengali (bn), Tamil (ta), Telugu (te), Malayalam (ml), Gujarati (gu), Punjabi (pa), and Odia (or):

• Studio Tier (onnx/*.onnx, 27.93 MiB): Full un-distilled 24,000 Hz HiFi-GAN v1 specialists (0.00 dB distillation loss, ~43 ms TTFA, 0.04x WebGPU RTF) automatically loaded on Desktop, WebGPU, and multi-core runtimes.

• Mobile Tier (onnx_mobile/*.onnx, 21.68 MiB): Warm-started 24,000 Hz c8c8_384 iSTFTNet specialists (10.91 GMAC/s, 4.1x fewer vocoder MACs, 0.66x real-device WASM RTF on Google Pixel 10 Pro XL, 519 ms first-chunk compute, 0 streaming underruns) automatically loaded on mobile browsers and low-memory CPUs.

• Web Studio (AudioWorklet + WebGPU / Mobile WASM SIMD): https://hemanth.github.io/ninada/

• GitHub Repository & Streaming SDKs (JS + Python): https://github.com/hemanth/ninada

• Curated 24 kHz Studio Split Manifests: https://huggingface.co/datasets/gnumanth/ninada-sanskrit-indic

What Is New in v5.1.2 (Clean Broadcast Studio Indian English + Dual-Tier HiFi-GAN v1 / c8c8_384 iSTFTNet)

• End-to-End Un-Frozen 48 kHz -> 24 kHz Studio Indian English (SPRINGLab/IndicTTS-English, CC-BY-4.0): Replaced donor-frozen English checkpoints with end-to-end un-frozen 24,000 Hz Studio Indian-English Male & Female specialists trained on clean single-speaker 48,000 Hz -> 24,000 Hz broadcast studio narrations (spec_flatness = 0.0005 Male / 0.0260 Female, 11x lower noise floor than v5.1.1; Whisper-small.en WER = 1.56% Male / 4.23% Female; Mobile iSTFTNet distillation MCD = 1.342 dB Male / 1.081 dB Female).
• Zero-Compromise Desktop Studio Quality + Real-Time Mobile CPU Streaming: Desktop and WebGPU clients run the full 27.93 MiB HiFi-GAN v1 24,000 Hz generator (onnx/*.onnx), while mobile and constrained CPU clients automatically load the distilled 21.68 MiB ISTFTWaveformDecoder (onnx_mobile/*.onnx, [8, 8] upsampling, 384 -> 192 -> 96 channels, n_fft = 16, hop = 4, total_hop = 256 samples) trained for 30,000 adversarial Multi-Period + Multi-Scale + Multi-Resolution STFT + Sub-Band steps (100% voiced across all 24 voices).
• Verified on Physical Mobile Hardware (Google Pixel 10 Pro XL, Chrome 156, 6 WASM SIMD Threads):
  • First-Chunk Compute (firstChunkPerfMs): 519 ms (861 ms audible start including Android hardware output latency)
  • Steady Mobile Vocoder RTF: 0.66x - 0.68x (~1.5x faster than real-time in pure browser WASM)
  • Streaming Continuity: 0 AudioWorklet underruns, 0 ms overlap, 35 ms natural inter-sentence breath, 56 - 61 dB multi-chunk streaming SNR vs monolithic synthesis

All 24 Published Specialists (v5.1.2, Studio onnx/ + Mobile onnx_mobile/)

Language	Studio / Mobile ONNX	Male Mobile MCD (0-8k)	Female Mobile MCD (0-8k)	8-12 kHz HF Err (M / F)	Voiced Rate (M / F)	Studio Corpus & License
Sanskrit (sa)	27.93 / 21.69 MiB	1.407 dB	1.382 dB	1.559 / 1.255 dB	100% / 100%	ai4bharat/Rasa (CC-BY-4.0)
Indian English (en)	27.92 / 21.68 MiB	1.342 dB	1.081 dB	5.178 / 1.352 dB	100% / 100%	SPRINGLab/IndicTTS-English (CC-BY-4.0)
Hindi (hi)	27.93 / 21.69 MiB	1.313 dB	1.178 dB	1.226 / 0.896 dB	100% / 100%	ai4bharat/Rasa (CC-BY-4.0)
Kannada (kn)	27.92 / 21.68 MiB	1.401 dB	1.329 dB	1.588 / 1.327 dB	100% / 100%	ai4bharat/Rasa (CC-BY-4.0)
Tamil (ta)	27.92 / 21.67 MiB	1.467 dB	1.125 dB	1.375 / 1.266 dB	100% / 100%	ai4bharat/Rasa (CC-BY-4.0)
Telugu (te)	27.93 / 21.68 MiB	1.447 dB	1.493 dB	1.377 / 1.363 dB	100% / 100%	ai4bharat/Rasa (CC-BY-4.0)
Malayalam (ml)	27.92 / 21.68 MiB	1.374 dB	1.336 dB	1.486 / 1.324 dB	100% / 100%	ai4bharat/Rasa (CC-BY-4.0)
Marathi (mr)	27.93 / 21.69 MiB	1.154 dB	1.314 dB	1.424 / 1.255 dB	100% / 100%	ai4bharat/Rasa (CC-BY-4.0)
Bengali (bn)	27.92 / 21.68 MiB	1.084 dB	1.767 dB	0.846 / 1.815 dB	100% / 100%	ai4bharat/Rasa (CC-BY-4.0)
Gujarati (gu)	27.93 / 21.69 MiB	1.182 dB	1.226 dB	0.923 / 0.957 dB	100% / 100%	ai4bharat/Rasa (CC-BY-4.0)
Punjabi (pa)	27.92 / 21.68 MiB	0.996 dB	1.214 dB	1.049 / 1.076 dB	100% / 100%	ai4bharat/Rasa (CC-BY-4.0)
Odia (or)	27.92 / 21.68 MiB	1.260 dB	2.092 dB	0.977 / 2.682 dB	100% / 100%	ai4bharat/Rasa (CC-BY-4.0)

Repository Layout

• onnx/{lang}.onnx & onnx/{lang}_female.onnx: All 24 full un-distilled 24,000 Hz Studio HiFi-GAN v1 W8A32+FP16 ONNX specialists (~27.93 MiB each) for Desktop, WebGPU, Python, and Node.js.
• onnx_mobile/{lang}.onnx & onnx_mobile/{lang}_female.onnx: All 24 distilled 24,000 Hz Mobile c8c8_384 iSTFTNet W8A32+FP16 ONNX specialists (~21.68 MiB each) for mobile browsers and constrained CPUs.
• checkpoints/{lang}_male.safetensors & checkpoints/{lang}_female.safetensors: All 24 full FP32 StudioVitsModel PyTorch checkpoints.
• vocoders_istft/{lang}_male_c8c8_384.safetensors & vocoders_istft/{lang}_female_c8c8_384.safetensors: All 24 standalone FP32 ISTFTWaveformDecoder checkpoints (29.85 MiB each).
• vocabs.json & models_manifest.json: Native-script tokenizers (12 scripts) and exact byte-verified dual-tier browser cache manifest.
• samples_v5/{lang}.wav & samples_v5/{lang}_female.wav: 24 reference 24,000 Hz showcase WAV clips.

24 kHz Streaming Quickstart

import ninada from 'ninada';

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

for await (const chunk of voice.stream('Welcome to Ninada, streaming twenty-four kilohertz studio speech in real time.')) {
  playPcm24k(chunk.pcm, chunk.sampleRate);
}

from ninada import Ninada

tts = Ninada(lang="sa", voice="female")
for chunk in tts.stream("नमस्ते, इदं निनाद चतुर्विंशतिसहस्रहर्ट्ज् ध्वनिपरीक्षणम् अस्ति॥"):
    play(chunk["pcm"], chunk["sample_rate"])  # 24,000 Hz float32 PCM

Attribution & Licensing

• Code & Runtime Engine: MIT License (hemanth/ninada).
• Model Weights & Training Corpora: Released under CC-BY-4.0, attributing AI4Bharat Rasa (ai4bharat/Rasa, CC-BY-4.0) for the 11 Indic languages (sa, hi, kn, ta, te, ml, mr, bn, gu, pa, or), IIT Madras SPRING Lab IndicTTS (SPRINGLab/IndicTTS-English, CC-BY-4.0) for Indian English (en Studio Male & Female), and IISc SPIRE Lab SYSPIN (CC-BY-4.0) initialization priors.