Abstract: Feature extraction and the size of dimension are important factors for implementing effective classification of heart sound signals. This paper presents a new method of extracting time-frequency features for heart sound signals based on adaptive cone-kernel. Firstly, time-frequency analysis normalized energy 10%E, 50%E, 90%E of heart sound signals and the largest energy frequency points Fmax are used to constitute the eight-dimensional feature vector of first and second heart sound signals. Secondly, the test parameters of Fisher dimension reduction method is adopted to conduct the comparative study for the correlation. Relatively small parameters F-10% and F-Fmax are selected to form the two-dimensional feature vector and apply them to the analysis of characteristic parameter divergence. The results show that the two-dimensional feature vector reflects the characteristics of heart sound signals comprehensively, improves the algorithm effectively and has some clinical values.