implement PCS

In the experiments, we used Short-time Fourier transform (STFT) with n_FFT=400, hop_length=100, windown_length=400, and Hamming window. Additionally, the audio we used in VoiceBank-DEMAND dataset was resampled to a sample rate of 16,000.

In the scenario mentioned above, the PCS we used is displayed on lines 10 ~ 20 of PCS400.py

Post-processing PCS

The proposed PCS can also be used as a post-processing (PP) method.
Here, we offer you an easy way to use it.

Requirements

• torch
• torchaudio
• librosa
• numpy
• scipy
• argparse

How to use ?

The audio to be processed should be placed in a folder. (denote as input_folder)

• You should modify the --input_folder (folder path), and --output_folder (folder path) in runPCS.sh first.
• Then execute sh runPCS.sh. The audio processed by PP-PCS will be saved to the --output_folder.

Notice that if the sample rate of input audio is not 16000, please modify the parameter of torchaudio.save() in PCS.py

Citation:

If you find the code useful in your research, please cite:

@article{chao2022perceptual,  
  title={Perceptual Contrast Stretching on Target Feature for Speech Enhancement},
  author={Chao, Rong and Yu, Cheng and Fu, Szu-Wei and Lu, Xugang and Tsao, Yu},
  journal={Proc. of INTERSPEECH},
  year={2022}
}