A repo for a solution to denoising and separating for two-speeker-mixed noisy speech, using a BSRNN inspired deep learning network.
View demos here.
| Key | VAlue |
|---|---|
| Datasets | AISHELL-3 & NoiseX-92 |
| FLOPs | 2.408G |
| Weights Size | 61.95M |
| Parameters | 16.15M |
Naive Case (only mix, no noise)
| Metric | SI-SNR | PESQ(wb) | PESQ(nb) | STOI |
|---|---|---|---|---|
| Raw dataset | 0.002 | 1.240 | 1.473 | 0.681 |
| BSRNN(modified) | 12.195 | 2.453 | 2.866 | 0.901 |
Difficult Case (with mix & noise)
| Metric | SI-SNR | PESQ(wb) | PESQ(nb) | STOI |
|---|---|---|---|---|
| Raw dataset | -0.597 | 1.146 | 1.379 | 0.656 |
| BSRNN(modified) | 11.384 | 2.212 | 2.661 | 0.880 |
Python version is 3.8. Requirements can be installed by:
pip install -r requirements.txt
See this if installing the dependency pesq fails.
Modify config/test.yml with your own dataset path, and run following command:
python speech-preprocess/test.py
Follow the order in data/index_data.py and data/make_data.py to config your raw dataset.
Then run following commands to make index and generate data:
python data/index_data.py
python data/make_data.py
For this project, AISHELL-3 and NoiseX-92 datasets are used.
Then, modify the config file config/train.yml and run following command:
python speech-preprocess/train.py
