liar1573/PAGAN

PAGAN: a phase-adapted GAN for speech enhancement

PAGAN

Dependencies

python3.0, pytorch=1.0.1, glob, librosa, numpy, pypesq, pystoi

Prepare Dataset

• Download Timit dataset
  Get Timit dataset at https://github.com/philipperemy/timit, contains a total of 6300 sentences, 630 speakers, we used these as clean dataset.
• Download noise dataset
  Get noise dataset at http://web.cse.ohiostate.edu/pnl/corpus/HuNonspeech/ HuCorpus.html or http://home.ustc.edu.cn/˜xuyong62/demo/115noises.html.
• Edit config.yaml

cd config
vim config.yaml

• preprocess ctc_data file

python data_for_1s.py
python data_rgan.py

Train/Test PAGAN

python train_rgan_ft_lps_multi_SN_real_ori.py
python test_PAGAN.py

Train/Test PAGAN_base

python train_PAGAN_base.py
python test_PAGAN_base.py

Train/Test rgan_ff

python train_rgan_ft_lps_multi_SN_real_ori_ff.py
python test_rgan_ff.py