# Audio Chat

## Running

```bash
uvicorn main:app --host 0.0.0.0 --port 8000
```

Client (for testing):
```bash
python client.py
```

## Architecture

Single-process FastAPI server. On each WebSocket connection a new `AudioSession` is created with three engines:

| Module | Purpose | Model (default) |
|--------|---------|-----------------|
| `engine/stt.py` | Speech-to-text | Systran/faster-whisper-large-v3 |
| `engine/llm.py` | LLM response generation | Qwen/Qwen2.5-7B-Instruct |
| `engine/tts.py` | Text-to-speech | facebook/mms-tts-rus |

Models are loaded lazily on first use if `initialize()` was not called. STT always runs in Russian (`language="ru"` with VAD).

## WebSocket Protocol

| Direction | Format | Meaning |
|-----------|--------|---------|
| Client → Server | `b"A" + PCM data` | Send audio chunk |
| Client → Server | `b"R"` | Reset conversation |
| Server → Client | `b"O" + WAV bytes` | LLM response as audio |
| Server → Client | `"TEXT:<transcription>"` | Recognized speech |

Audio format: 16-bit PCM mono, 16 kHz input / 24 kHz output.

## Configuration

All settings via `.env` (loaded by `config.py`). Key vars:

- `DEVICE` — `"cuda"` or `"cpu"` (default `"auto"`)
- `AUDIO_BUFFER_SECONDS` / `CHUNK_SIZE` — silence detection thresholds
- `LLM_MAX_TOKENS` / `LLM_TEMPERATURE` — generation parameters

## Gotchas

- No test suite or linting configured.
- Models download on first use; ensure network access to HuggingFace.
- `AudioSession` holds conversation history (last 6 turns) in memory — each WebSocket reconnect resets it.
- Thread pool executor is fixed at 2 workers; concurrent heavy requests will queue.
- TTS pipeline falls back to CPU (`device=-1`) if GPU initialization fails silently.