Nonlinear Schur-Type Audio Signal Parameterization for Convolutional Networks

Abstract

This article introduces a novel signal parameterization approach, termed nonlinear Schur-type signal parameterization, designed to enhance machine learning tasks such as signal classification and recognition. Traditional linear parameterization methods often struggle with the complex, nonlinear nature of real-world data. The mathematical foundation of the proposed parameterization method is extraction of Schur coefficients. The presented method is scalable and can be adjusted to the signal nature. The nonlinear Schur parameterization produces a matrix of Schur coefficients in time, dedicated to be a 2D input of convolutional neural networks (CNN). The performed experiments for the audio signals from open access datasets show that the signal representation in the form of the Schur coefficients is very efficient for recognition performance. The results obtained by CNN show an improvement in the classification accuracy in comparison with solutions based on preprocessing in frequency domain as FFT or MFCC.