ABSTRACT
ABSTRACT: Recently Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) were studied and incorporated into implementation of snore detection system in mobile phone environment. In this manuscript modified methods for automatic snore detection, inspired by existing scientific articles, are presented. The main goal of these modifications is to increase the robustness of snore/non-snore signal classification. First the weak spots of each solution are analyzed. For both PCA and LDA analysis, features were computed from output of Short Time Fourier Transform (STFT)—a square sum of amplitudes in given equally sized sub-band (15 × 500 Hz) of the frequency range 0-7500 Hz at sampling frequency equal to 16 kHz. PCA and LDA based detection methods were simulated and compared. 15-dimensional features are calculated and projected into two-dimensional classification sub-space. The results of the comparison showed that due to different behavior of PCA and LDA, utilization of different classification methods is required. Following the knowledge obtained from the analysis one implementation is proposed and described.
