Separation of PCG signal from Mixture of Speech and PCG Signals with Genetic Algorithm-Based Filter Banks

The aim of the paper is to separate the phonocardiographic (PCG) signal from the mixture of PCG and speech signals. Therefore, genetic algorithm (GA) based filter-banks approach has been used to separate the signal. In this proposed technique, speech signal was modified and then subtracted from mixed signal to obtain the PCG signal. The modification in the speech was performed by modifying the short-time Fourier transform magnitude response. The magnitude response was further decomposed into nine filter-banks, each of bandwidth 50 Hz, upto 450 Hz. The magnitude components in each filter-band were varied with the weights of GA. The phase component was not modified. Extracted PCG signals were evaluated using Mel-frequency cepstral coefficients (MFCCs) based Mahalanobis distance measure and perceptual evaluation of speech quality. It is observed that GA performs well to extract PCG signal form mixture, with population size 30, number of weights 10 and the number of iterations less than 80. The proposed technique shows better accuracy than FastICA, however, the main limitation of the GA-based filter banks is the time complexity arising due to the involvement of iterative behavior.