Acoustic sound generated by the heart mechanical activity, can provide useful information about the condition of heart valves. The heart sound auscultation is the fundamental tool in the evaluation of the cardiovascular system. The advantage of this method is fast, inexpensive and noninvasive. Due to human auscultatory limitation and non-stationary characteristics of phonocardiogram signals (PCG), diagnosis based on sounds that are heard via a stethoscope is difficult skill, therefor it requires a lot of practice. This study has proposed a biomedical automatic system for classification of PCG signals, which, recorded by a digital stethoscope. In order to extract various characteristics of PCG signals, the power spectrum estimation, wavelet transform (WT) and Mel frequency Cepstrum coefficients (MFCC) have been used in feature extraction step. Features are given to four classifiers: support vector machine (SVM), k-nearest neighbor (k-NN), multilayer perceptron (MLP) and maximum likelihood (ML). The majority voting combination rule is utilized for fusion of different classifiers. The proposed method has been examined on dataset of 90 PCG records containing healthy and three types of cardiac valve diseases (pulmonary stenosis (PS), Atrial Septal Defect (ASD) and Ventricular Septal Defect (VSD)). The experimental results demonstrate that the classifier fusion rule significantly increases the diagnostic accuracy of abnormal PCG. Our proposed method can be used for online classification of PCG in intelligent diagnosis systems.