swanada (स्वनद)

In-browser, Node.js, and Python 16,000 Hz multilingual speech-to-text (STT) for Classical Sanskrit (sa), Indian English (en), and ten Indic languages (hi, kn, mr, bn, ta, te, ml, gu, pa, or) powered by 17-layer Conformer-CTC INT8 models running 100% client-side via ONNX/WASM with zero cloud or server dependencies.

npm install swanada

pip install swanada

Quick Start (JavaScript / TypeScript)

import { Swanada } from 'swanada';

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

// Transcribe audio Float32Array PCM or WAV ArrayBuffer
const result = await transcriber.transcribe(audioPcmFloat32);
console.log(`Transcript (${result.language}):`, result.text);
console.log(`Duration: ${result.durationMs}ms | Latency: ${result.inferenceMs}ms | RTF: ${result.rtf}`);

// Classical Sanskrit to Roman IAST transliteration
const iast = Swanada.transliterateIast(result.text);
console.log('IAST Transliteration:', iast);

await transcriber.release();

Quick Start (Python)

from swanada import Swanada

stt = Swanada.create(language="kn")  # Kannada
result = stt.transcribe("speech_16k.wav")

print("Transcript:", result.text)
print(f"Duration: {result.duration_ms:.1f}ms | Latency: {result.inference_ms:.1f}ms | RTF: {result.rtf:.3f}")

CLI Usage

Dual-runtime command-line interface with identical flags:

# Node.js CLI
npx swanada speech.wav -l sa
npx swanada recitation.wav -l sa --json -o output.json
cat stream.wav | npx swanada -l hi --stream

# Python CLI
swanada speech.wav -l kn
cat stream.wav | swanada -l ta --stream

Unified CLI Flags

• -l, --language <lang>: Language code (default: sa). Supported: sa, en, hi, kn, mr, bn, ta, te, ml, gu, pa, or.
• -m, --model-dir <dir>: Optional custom models directory.
• -o, --output <file>: Output text file (default: stdout).
• --json: Emit structured JSON with text, durationMs, inferenceMs, rtf, and tokens.
• --stream: Simulate chunk-by-chunk streaming transcription (320 ms quanta).
• --iast: Transliterate Classical Sanskrit output to Roman IAST with Vedic accents.
• -h, --help: Show help and usage summary.
• -v, --version: Show version (swanada 1.0.2).

Benchmarks

Evaluated across 480 held-out 16 kHz utterances (40 per language, B = 1,000 bootstrap 95% CIs), 1,200 unseen google/fleurs + wild test clips, 150 real-world multi-accent / conversational slang utterances (MushanW/GLOBE, AMI, regional accents, Indic loanwords), and 72 streaming sessions (100 ms chunks):

Conversational Slang, Regional Accents & Reverberant Mic Suite (n = 150, B = 1,000 Bootstrap)

Shared 178 MB INT8 Conformer encoder + 515 KB per-language CTC projection heads run at 0.0325x CPU RTF, ~8-11 ms streaming chunk latency, and 0.00% committed prefix churn.

Live Playground

• h3manth.com AI Experiments Studio: https://h3manth.com/ai/swanada/
• GitHub Pages Web Studio: https://hemanth.github.io/swanada/

Model Weights & Hugging Face Hub

Quantized INT8 model weights are hosted on Hugging Face Hub:

• Repository: https://huggingface.co/gnumanth/swanada

License

MIT © Hemanth HM