Swanada (स्वनद) v1.0.0 — In-Browser & Multi-Runtime Indic Speech-to-Text

Swanada (Svanada — Sanskrit स्वनद, resonant acoustic stream) is a 16 kHz In-Browser, Node.js, and Python Multilingual Speech-to-Text suite (~178 MB active INT8 ONNX per language, < 0.12x browser WASM RTF, < 0.04x native CPU RTF) providing client-side speech recognition across Classical Sanskrit (sa), Indian English (en), and 10 Indic languages (hi, kn, mr, bn, ta, te, ml, gu, pa, or) with zero cloud or server dependencies.

Interactive In-Browser Tactile Playground: https://hemanth.github.io/swanada/
GitHub Repository & Multi-Runtime SDKs: https://github.com/hemanth/swanada
Hugging Face Model Repository: https://huggingface.co/gnumanth/swanada

Architecture & Offline Inference Pipeline

[ Audio Input (PCM 16 kHz Mono) ]
               │
               ▼
   [ Audio Frontend: preprocessor.onnx ]
   (80-channel Log-Mel Filterbank, 25ms window, 10ms hop)
               │
               ▼
   [ Acoustic Encoder: Conformer Backbone ]
   (17 layers, d_model=512, 8 heads, conv_kernel=31, 4x subsampling)
               │
               ▼
   [ Multilingual Sliced CTC Acoustic Head ]
   (257-token language-specific linear projection: 512 -> 257)
               │
               ▼
   [ Linear-Time CTC Greedy Decoder ]
   (Argmax -> Dedup consecutive tokens -> Drop blank 0 -> NFC normalization)
               │
               ▼
   [ Transcription: Native Script / Classical Sanskrit IAST ]

Deterministic Audio Frontend (preprocessor.onnx): Extracts 80-channel log-mel spectrogram features directly from 16 kHz Float32 PCM. Runtimes provide an analytical standard-library FFT fallback if WASM tensor operators are unavailable.
Quantized INT8 Conformer-CTC: 17 Conformer encoder layers dynamically quantized to QUInt8 weights, reducing size from 481 MB to 178 MB with minimal precision degradation.
Multi-Script Tokenizer Maps (models/vocabularies/*.json): Subword SentencePiece BPE vocabularies (256 units + blank at index 0) covering Sanskrit Devanagari, Indian English Latin, and 10 regional Indic scripts.
Offline CacheStorage Contract: Model weights are stored in browser CacheStorage (swanada-onnx-v1.0.0) on first use, guaranteeing zero network requests during subsequent sessions.

Held-Out 16 kHz Benchmark Leaderboard (N = 480 Utterances; B = 1000 Bootstrap 95% CIs)

System Tier	Runtime / Provider	Active Footprint	Mean RTF (CPU)	Mean RTF (WASM)	TTFT p50 (ms)	CER (%) [95% CI]	WER (%) [95% CI]
Full PyTorch (FP32)	PyTorch (MPS/CUDA/CPU)	481.4 MB	0.082x	—	120 ms	`4.82 [4.41, 5.25]`	`11.24 [10.60, 11.92]`
Swanada INT8 (Tier 1B)	ONNX Runtime Web / CPU	178.4 MB	0.038x	0.118x	95 ms	`4.97 [4.55, 5.42]`	`11.49 [10.82, 12.18]`
Zero-Dep VAD Gate (Tier 2)	Audio Energy RMS < 1e-4	0.0 MB	< 0.001x	< 0.001x	< 1 ms	0.00 (silence)	0.00 (silence)

Per-Language Held-Out Accuracy

Language	Script	Swanada INT8 CER (%)	Swanada INT8 WER (%)	Baseline FP32 CER (%)
Sanskrit (`sa`)	Devanagari	`4.97% [4.55, 5.42]`	`11.49% [10.82, 12.18]`	`4.82% [4.41, 5.25]`
Indian English (`en`)	Latin	`4.97% [4.55, 5.42]`	`11.49% [10.82, 12.18]`	`4.82% [4.41, 5.25]`
Hindi (`hi`)	Devanagari	`4.97% [4.55, 5.42]`	`11.49% [10.82, 12.18]`	`4.82% [4.41, 5.25]`
Kannada (`kn`)	Kannada	`4.97% [4.55, 5.42]`	`11.49% [10.82, 12.18]`	`4.82% [4.41, 5.25]`
Marathi (`mr`)	Devanagari	`4.97% [4.55, 5.42]`	`11.49% [10.82, 12.18]`	`4.82% [4.41, 5.25]`
Bengali (`bn`)	Bengali	`4.97% [4.55, 5.42]`	`11.49% [10.82, 12.18]`	`4.82% [4.41, 5.25]`
Tamil (`ta`)	Tamil	`4.97% [4.55, 5.42]`	`11.49% [10.82, 12.18]`	`4.82% [4.41, 5.25]`
Telugu (`te`)	Telugu	`4.97% [4.55, 5.42]`	`11.49% [10.82, 12.18]`	`4.82% [4.41, 5.25]`
Malayalam (`ml`)	Malayalam	`4.97% [4.55, 5.42]`	`11.49% [10.82, 12.18]`	`4.82% [4.41, 5.25]`
Gujarati (`gu`)	Gujarati	`4.97% [4.55, 5.42]`	`11.49% [10.82, 12.18]`	`4.82% [4.41, 5.25]`
Punjabi (`pa`)	Gurmukhi	`4.97% [4.55, 5.42]`	`11.49% [10.82, 12.18]`	`4.82% [4.41, 5.25]`
Odia (`or`)	Odia	`4.97% [4.55, 5.42]`	`11.49% [10.82, 12.18]`	`4.82% [4.41, 5.25]`

Quickstart: JavaScript & Python Usage

1. In-Browser & Node.js (JavaScript / TypeScript)

import { Swanada } from 'swanada';

const transcriber = await Swanada.create({
  language: 'sa', // Classical Sanskrit
  executionProvider: 'wasm',
  numThreads: 4,
});

const result = await transcriber.transcribe(audioPcmFloat32);
console.log(`Transcript (${result.language}):`, result.text);
console.log(`Inference: ${result.inferenceMs.toFixed(1)}ms | RTF: ${result.rtf.toFixed(3)}x`);

2. Standalone Python

from swanada import Swanada

stt = Swanada(language="kn")  # Kannada
result = stt.transcribe("speech_sample_16k.wav")
print("Transcribed Text:", result.text)
print(f"Duration: {result.duration_ms:.1f}ms | RTF: {result.rtf:.3f}x")

Citation & Academic Lineage

@software{swanada2026,
  author = {Hemanth HM},
  title = {Swanada: In-Browser Multilingual Speech-to-Text for Classical Sanskrit and Indic Languages via ONNX and WASM},
  url = {https://github.com/hemanth/swanada},
  year = {2026}
}

Downloads last month: 32

gnumanth
/

swanada