# Alibaba Cloud DashScope ASR — Reference Examples

Adapted from `/mnt/c/Users/woody/Documents/projects/voice input/` (Cantonese voice-to-text web app).

## Files

| File | What | Language |
|------|------|----------|
| `alibaba_asr_backend.py` | FastAPI WebSocket proxy to DashScope realtime ASR | Python |
| `alibaba_asr_frontend_vanilla.html` | Browser audio capture + WebSocket (vanilla JS) | HTML/JS |
| `alibaba_asr_frontend_react.tsx` | React/TS hook + component for audio capture | TypeScript/React |

## Architecture

```
Browser (Float32 PCM, 16kHz mono)
  │  WebSocket: send(float32Data.buffer)
  ▼
FastAPI Backend (/ws/asr/{video_id})
  │  Convert Float32 → S16_LE → base64
  ▼
Alibaba Cloud DashScope (wss://dashscope-intl.aliyuncs.com/api-ws/v1/realtime)
  │  Model: qwen3-asr-flash-realtime
  ▼  Language: yue (Cantonese)
Transcript JSON → Browser
```

## Key Details

- **Audio format**: Float32 PCM, 16kHz, mono (browser) → S16_LE PCM, 16kHz, mono, base64 (DashScope)
- **Model**: `qwen3-asr-flash-realtime` (WebSocket realtime, unlimited duration)
- **Endpoint**: `wss://dashscope-intl.aliyuncs.com/api-ws/v1/realtime`
- **SDK**: `pip install dashscope>=0.4.0`
- **Cantonese**: Language code `yue` (works natively with DashScope)
- **VAD**: Server-side (Alibaba Cloud handles voice activity detection)
- **Pricing**: ~$0.00009/second
- **Features**: Punctuation, ITN, filler word filtering, multi-language auto-detect

## Dependencies

```
# Python
dashscope>=0.4.0
openai>=1.52.0
zhconv>=1.4.0       # Simplified → Traditional Chinese (optional)

# No additional JS deps needed — native Web APIs only:
#   WebSocket, AudioContext, ScriptProcessorNode, getUserMedia
```