A small, PyTorch-free runtime for speech enhancement based on LavaSR.

This project keeps the original LavaSR idea and model structure, but runs inference with ONNX Runtime + NumPy/SciPy + SoundFile instead of PyTorch.

This app is derived from the original ysharma3501/LavaSR repository.

From the original repository, LavaSR is a lightweight speech enhancement / restoration system for improving low-quality speech. The upstream project describes support for speech bandwidth extension up to 48 kHz, optional denoising, and use cases such as TTS enhancement, real-time enhancement, and dataset restoration. The repository also explains that the method uses a Vocos-based architecture for bandwidth extension together with a Linkwitz-Riley-inspired refiner, and acknowledges UL-UNAS for the denoiser component.

This runtime-only version focuses on inference only. Model export/conversion is out of scope here.

This app runs the pipeline with:

• ONNX Runtime for the denoiser core and enhancer neural networks
• NumPy / SciPy for STFT, ISTFT, resampling, mel features, and spectral-domain glue logic
• SoundFile for WAV/FLAC audio I/O

The current inference flow is:

1. Load audio at 16 kHz
2. Optionally denoise with the ONNX ULUNAS-derived denoiser core
3. Resample to 48 kHz
4. Run the ONNX enhancer backbone and ONNX spectrogram head
5. Reconstruct waveform and apply the final spectral merge

Expected runtime file structure:

.
├── main.py
├── lavasr_core.py
├── config.yaml
├── denoiser_core_legacy_fixed63.onnx
├── enhancer_backbone.onnx
├── enhancer_backbone.onnx.data
├── enhancer_spec_head.onnx
└── enhancer_spec_head.onnx.data

The ONNX files are hosted in the release assets here:

• Release page: https://github.com/Topping1/LavaSR-ONNX/releases/tag/Alpha-v0.1
• denoiser_core_legacy_fixed63.onnx: https://github.com/Topping1/LavaSR-ONNX/releases/download/Alpha-v0.1/denoiser_core_legacy_fixed63.onnx
• enhancer_backbone.onnx: https://github.com/Topping1/LavaSR-ONNX/releases/download/Alpha-v0.1/enhancer_backbone.onnx
• enhancer_backbone.onnx.data: https://github.com/Topping1/LavaSR-ONNX/releases/download/Alpha-v0.1/enhancer_backbone.onnx.data
• enhancer_spec_head.onnx: https://github.com/Topping1/LavaSR-ONNX/releases/download/Alpha-v0.1/enhancer_spec_head.onnx
• enhancer_spec_head.onnx.data: https://github.com/Topping1/LavaSR-ONNX/releases/download/Alpha-v0.1/enhancer_spec_head.onnx.data

Install the runtime dependencies with:

pip install numpy scipy soundfile onnxruntime

For GPU inference, use the ONNX Runtime package appropriate for your platform instead of the default CPU package.

Basic run:

python main.py input.wav -o output.wav

Run with denoising enabled:

python main.py input.wav -o output.wav --denoise

• This repository is intended for runtime inference.
• It does not require PyTorch for normal execution.
• The denoiser ONNX model is used with fixed-size internal chunks in the runtime code.
• The original LavaSR project remains the reference implementation for the model and overall method.

• Original project: ysharma3501/LavaSR