Analyzing the Use of Auditory Filter Models for Making Interpretable Convolutional Neural Networks for Speaker Identification

Abstract

Most of the advances in artificial intelligence are based on understanding the function of different parts of the living organism's body. High complexity of some body parts, e.g., the brain, leads to using some abstractions for making intelligent models, which can make the models uninterpretable. This general process can be seen in the development of Deep Neural Networks (DNNs). Although DNNs are models with high performance, they have a black-box nature which makes them unreliable in some applications such as medicine. Fortunately, nature can again help to make DNN models explainable. The use of meaningful filters in the first layer of Convolutional Neural Networks (CNNs) is a successful attempt in this field. The goal of this paper is to examine the use of some auditory filter models as CNN front-ends to make them interpretable and then to evaluate the resulting filter banks in the Speaker Identification (SID) task. Results confirm the previous knowledge about the filtering mechanism of the auditory system. This simple observation can lead to an abstract conclusion that making a complex learning model interpretable, specifically using simple elements inspired by nature, can disclose the hidden aspects of how the human body works. Moreover, replicating the essential functions of the human auditory system leads to better model performance.