# Local STT (Speech-to-Text) SOP

## Prerequisites
- `pip install openai-whisper`
- Ensure `ffmpeg` is installed (`brew install ffmpeg`).

## Usage
```python
import whisper
model = whisper.load_model("base") # Use 'base' or 'tiny' for speed on Apple Silicon
result = model.transcribe("audio.wav")
print(result["text"])
```

## Audio Capture (R14 Verified)
- **Microphone**: Verified with `sounddevice`. Use `sd.query_devices()` to identify input index.
- **System Loopback (Internal Audio)**:
  - macOS **does not** support system audio capture natively. Standard APIs (`avfoundation`, `coreaudio`) only see physical hardware.
  - **Requirement**: Virtual driver **BlackHole** (2ch or 16ch) is mandatory for loopback.
  - **Workflow**: Install BlackHole -> Set macOS Audio Output to "BlackHole" -> Select "BlackHole" as Input Device in `sounddevice`.

## Notes
- Running large models (`medium`, `large`) on Apple Silicon CPU is slow. Use `faster-whisper` for production.