PGUSE: A Composite Predictive-Generative Approach to Monaural Universal Speech Enhancement

This is the official implementation of PGUSE model.

Installation

1. Prepare a virtual environment with python, pytorch, and pytorch_lightning. (We use python==3.10.14, pytorch==2.0.0, and pytorch_lightning==2.0.7, but other versions probably also work.)
2. Install the package dependencies via pip install -r requirements.txt.

Training

Before training, please check config.yaml to set hyperparameters, including devices, logdir, dataset path, ...

Then you can train the model by:

python train.py --config ./config/config.yaml

Testing

First, specify the ckpt_path in config/config.yaml, and then run:

python test.py --config ./config/config.yaml --save_enhanced <path-to-savedir>

Reference

sgmse

sgmse-bbed

LiSenNet