CVD (Cardiovascular disease) has become a major threat to national health, and heart sound classification and recognition help to diagnose CVD. To improve the accuracy of heart sound signal recognition, an improved MFCC (Mel-scale Frequency Cepstral Coefficients) method is proposed. The method is based on the MoDWT (Maximum Overlap Discrete Wavelet Transform) to de-decompose the signal and aims to enhance the extraction of data features. First, the open-source heart sound dataset was preprocessed and MoDWT was used to decompose the signal. Then, time-domain features are extracted using the improved MFCC method. Finally, the fused features are classified and recognized. The experimental results verify that the binary classification experiments on the PhysioNet 2016 public dataset achieved an accuracy of 96.1%, a recall of 95.3%, a precision of 100.0%, a specificity of 100.0%, and an F1-score of 97.6%. Experimental results show that this method can effectively improve the recognition rate of heart sound classification.