A performance evaluation of several artificial neural networks for mapping speech spectrum parameters

Abstract

In this work, we compare different neural network architectures, for the task of mapping spectral coefficients of noisy speech signals with those corresponding to natural speech. In previous works on the subject, fully-connected multilayer perception (MLP) networks and recurrent neural networks (LSTM & BLSTM) have been used. Several references report some initial trial and error processes to determine which architecture to use. Finding the best network type and size is of great importance due to the considerable training time required by some models of recurrent networks. In our work, we conducted extensive tests training more than five hundred networks, with several architectures to determine which cases present significant differences. The results show that for this application of neural networks, the architectures with more layers or the greater number of neurons are not the most convenient, both for the time required in their training and for the adjustment achieved. These results depend on the complexity of the task (the signal-to-noise ratio or SNR) and the amount of data available. This exploration can guide the most efficient use of these types of neural networks in future mapping applications, and can help to optimize resources in future studies by reducing computational time and complexity.