# LavaSR — bundled mirror (v2)

Self-contained inference bundle for the MAESTRO desktop app. One ~55 MB download → fully offline 48 kHz speech enhancement.

## Contents

- `enhancer_v2/model.safetensors` — Vocos-based bandwidth-extension weights (converted from upstream `pytorch_model.bin` via `safetensors.torch.save_model` for shared-buffer safety)
- `enhancer_v2/config.yaml` — Vocos hyperparameters (verbatim from upstream)
- `denoiser/denoiser.safetensors` — UL-UNAS denoiser weights (converted from upstream `denoiser.bin`)

## Source

- Enhancer + denoiser bundle: [YatharthS/LavaSR](https://huggingface.co/YatharthS/LavaSR)
- LavaSR code: [github.com/ysharma3501/LavaSR](https://github.com/ysharma3501/LavaSR) — Vocos backbone adapted for speech BWE, ~5000× realtime on GPU.
- UL-UNAS denoiser: [github.com/Xiaobin-Rong/ul-unas](https://github.com/Xiaobin-Rong/ul-unas).

## License

- LavaSR (enhancer): Apache-2.0
- UL-UNAS (denoiser): MIT (Xiaobin-Rong)

Both upstreams permit commercial redistribution with attribution.