Objective To introduce the least square support vector machine (LS-SVM) into the recognition of heart sound, as well as optimizing its parameters setting to obtain the optimal classification results. Methods 99 heart sounds were obtained from our hospital and the internet. Two samples of 5 s were extracted from each heart sound to construct one training set and two test sets. 3-layer wavelet packets decomposition of sym6 was applied to each sample to extract feature. Then, the training set was used to machine learning of SVM and LS-SVM. One test set was used to parameters optimization, the other was for test of optimized SVM and LS-SVM. Results The C and σ of the SVM that examined by Gaussian radial basis function were both 20.086. The accuracy for first test set was the highest (79.7%). For second test set, the accuracy was 84.5%, and the running times were 0.108 s and 0.117 s, respectively. For the LS-SVM, the accuracy for first test set was the highest (94.2%) while σ2=1 and γ=20.086. For second test set, the accuracy was 89.6%, and the running times were 0.0638 s and 0.0692 s, respectively. Conclusion The LS-SVM that find local optimal solution based on the linear equation method can operate faster, and it is more suitable for recognition of heart sound samples.