LPS_extraction

============================================

The script is to extract log-power-spectrum (LPS) for speech enhancement and bandwidth extension.

Requirements

The model is implemented in PyTorch and uses several additional libraries. Specifically, we used:

• pytorch==1.0
• python==3.6.8
• numpy==1.15.4
• scipy==1.2.0

Setup

To install this package, simply clone the git repo:

git clone https://github.com/nanahou/LPS_extraction.git;
cd LPS_extraction;

Contents

The repository is structured as follows.

• ./data: some audio samples from dataset[1]
• audioread.py: the function to read audios
• extract_LPS.py: the main scripts to extract features
• normhamming.py: the function to apply a normalized square root hamming periodic window
• plot_spectrum.py: the function to plot the LPS features
• sigproc.py: including the functions to frame signals, deframe signals from [2]

Usage

• If extracting LPS features, you only need to replace the path in extract_LPS.py with your own data path and run:

  python extract_LPS.py

• If plotting your features, you only need to call the function in plot_spectrum.py.

Reference

[1]. Valentini-Botinhao, C., Wang, X., Takaki, S. and Yamagishi, J., 2016. Speech Enhancement for a Noise-Robust Text-to-Speech Synthesis System Using Deep Recurrent Neural Networks. In Interspeech (pp. 352-356).
[2]. https://github.com/jameslyons/python_speech_features/blob/master/python_speech_features/sigproc.py