Gender Detection with Heart Sound Using MFCC-based Statistical Features

This paper investigates the utility of various MFCC (Mel frequency cepstral coefficients)-based statistical features extracted from heart sound signal for the purpose of gender detection. The proposed method consists of three steps as feature extraction, support vector machine (SVM) training and testing unknown subjects. First, MFCC features are extracted from heart sounds, their various statistics are calculated and combined together to construct a new augmented features vector. Second, the statistical models based on SVM are trained by these new feature vectors. Finally, the unknown subject is tested by the proposed system to infer a decision about his or her gender. Experiments have been evaluated on a publicly available database from 98 (49 females and 49 males) subjects. The results show that the use of new combined feature vectors increased average correct classification rate from 91.83 to 93.87.