The Power Cepstrum Calculation with Convolutional Neural Networks

Authors

  • Mario Alejandro García Universidad Tecnológica Nacional
  • Eduardo Atilio Destéfanis Universidad Tecnológica Nacional Facultad Regional Córdoba, Argentina

DOI:

https://doi.org/10.24215/16666038.19.e13

Keywords:

Cepstrum, Discrete Fourier transform, Spectrogram, Deep learning, Convolutional neural network

Abstract

A model of neural network with convolutional layers that calculates the power cepstrum of the input signal is proposed. To achieve it, the network calculates the discrete-time short-term Fourier transform internally, obtaining the spectrogram of the signal as an intermediate step. The weights of the neural network can be calculated in a direct way or they can be obtained through training with the gradient descent method. The behaviour of the training is analysed. The model originally proposed cannot be trained in a complete way, but both the part that calculates the spectrogram and also a variant of the cepstrum equivalent to the autocovariance that keeps a big part of its usefulness can be trained. For the cases of successful training, an analysis of the obtained weights is done. The main conclusions indicate, on the one hand, that it is possible to calculate the power cepstrum with a neural network; on the other hand, that it is possible to use these networks as the initial layers of a deep learning model for the case of trainable models. In these layers, weights are initialised with the discrete Fourier transform (DFT) coefficients and they are trained to adapt to specific classification or regression problems.

Downloads

Download data is not yet available.

Downloads

Published

2019-10-10

How to Cite

García, M. A., & Destéfanis, E. A. (2019). The Power Cepstrum Calculation with Convolutional Neural Networks. Journal of Computer Science and Technology, 19(2), e13. https://doi.org/10.24215/16666038.19.e13

Issue

Section

Original Articles

Most read articles by the same author(s)