# Inflect-Micro-v2 TTS SDK for AX650 NPU3

Encoder-Decoder split VITS text-to-speech engine.

## Files
- `inflect_encoder.axmodel` — Text encoder (NPU)
- `inflect_decoder.axmodel` — Flow + HiFi-GAN decoder (NPU)  
- `model.pth` — Full PyTorch weights (CPU components: embedding, duration, flow params)
- `config.json` — Model configuration
- `tts_engine.py` — Main inference engine

## Usage
```python
from inflect_tts_sdk import InflectTTSEngine

engine = InflectTTSEngine()
engine.warmup()  # Initialize NPU sessions

sr, wav = engine.synthesize("Hello world!")
engine.save("Hello world!", "output.wav")
```

## Pipeline
1. Text → Phonemes → Tokens (CPU, Python frontend)
2. Embedding lookup (CPU)
3. **Encoder NPU**: embeddings → m_p, logs_p, x_mask
4. Duration predictor + alignment (CPU)
5. **Decoder NPU**: z_p → waveform

## Requirements
See `requirements.txt`.