Power Cepstrum Calculation with Convolutional Neural Networks.

Thumbnail Image

Date

2019

Journal Title

Journal ISSN

Volume Title

Publisher

Journal of Computer Science y Technology.

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 interme diate step. Although the proposed neural networks weights can be calculated in a direct way, it is nec essary to determine if they can be obtained through training with the gradient descent method. In order to analyse the training behaviour, tests are made on the proposed model, as well as on two variants (power spectrum and autocovariance). Results show that the calculation model of power cepstrum cannot be trained, but the analysed variants in fact can.
Se propone un modelo de red neuronal con capas de convolución que calcula el power cepstrum de una señal de audio. Para logarlo, la red calcula internamente la Transformada Discreta de Fourier de Tiempo Reducido, obteniendo el espectrograma de la señal como paso intermedio. Si bien los pesos de la red neuronal propuesta se pueden calcular de forma directa, uno de los objetivos de este trabajo es determinar si esta puede ser entrenada con el me´todo del gradiente descendiente. Para analizar el comportamiento del entrenamiento se realizan pruebas sobre el modelo propuesto y también sobre dos variantes (power spectrum y auto covarianza). Los resultados indican que el modelo de calculo del power cepstrum no se puede entrenar, pero las variantes analizadas s´ı.

Description

Keywords

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

Citation

Journal of Computer Science y Technology, 2019.

Endorsement

Review

Supplemented By

Referenced By

Creative Commons license

Except where otherwised noted, this item's license is described as info:eu-repo/semantics/openAccess