# Stream ASR — 100% On-Device Streaming Speech Recognition

> Ultra-fast, browser-native WebAssembly streaming speech recognition runtime powered by MNN & SIMD128. Zero cloud dependencies, zero audio uploaded to any server, fully offline-ready.

## About
- Site: https://asr.firetable.tech
- GitHub: https://github.com/firetable/stream-asr
- Package: `@firetable/stream-asr`
- Author: FireTable
- License: Apache-2.0
- Architecture: WebAssembly + SIMD128 + Web Workers + AudioWorklet

## Key Capabilities
- **Sub-50ms Chunk Latency**: Real-time acoustic processing with immediate text output as the user speaks.
- **100% Client-Side Privacy**: Zero audio bytes uploaded to the cloud. All computation occurs locally on the user's hardware.
- **Full Offline Operation**: Browser IndexedDB / CacheStorage caches model tensors and WASM binaries for instant subsequent boots without network access.
- **Hardware Acceleration**: SIMD128-vectorized MNN tensor kernels delivering RTF 0.45x on standard single-core CPUs.
- **Lightweight Footprint**: ~175MB quantized model weights with < 250MB peak memory during live recognition.
- **Languages Supported**: Chinese (Mandarin) and English bilingual recognition.

## Quick Start (TypeScript / JavaScript)
```bash
# Install via your project's package manager
npm install @firetable/stream-asr
# or: pnpm add @firetable/stream-asr
```

```typescript
import { StreamRecognizer } from '@firetable/stream-asr';

const recognizer = new StreamRecognizer({
  modelUrl: '/model.bin',
  wasmBaseUrl: '/wasm',
  mode: 3, // Universal Streaming + Neural VAD
  vadThreshold: 0.5,
  enableCache: true,
});

recognizer.on('download_progress', (e) => {
  console.log(`Downloading: ${e.percent}% (${e.loaded}/${e.total} bytes)`);
});

recognizer.on('ready', () => {
  console.log('Engine ready for live audio streaming');
});

recognizer.on('recognizing', ({ text, vadState }) => {
  console.log('[Streaming Partial]:', text);
});

recognizer.on('recognized', ({ text, tokenCount, inferenceMs }) => {
  console.log('[Sentence Final]:', text);
});

await recognizer.init();

// Stream 16kHz Float32 PCM chunks from Web Audio API
// (Built-in auto-resampler supports 44.1kHz / 48kHz seamlessly)
recognizer.feed(pcmFloat32Array, audioContext.sampleRate);
```

## System Requirements
- Modern Chromium-based browser (Chrome, Edge 90+)
- Safari 16.4+ (iOS & macOS)
- Firefox 90+
- WebAssembly with SIMD128 support (zero COOP/COEP isolation headers needed)

## Resources & Links
- [Interactive Studio](https://asr.firetable.tech/#hero)
- [Architecture & Benchmarks](https://asr.firetable.tech/#features)
- [Quick Start Guide](https://asr.firetable.tech/#cta)
- [Full LLM Knowledge Base](https://asr.firetable.tech/llms-full.txt)
- [Agent Auth Policy](https://asr.firetable.tech/auth.md)