In this paper, a new technique for human identification task based on heart sound signals has been proposed. It utilizes a feature level fusion technique based on canonical correlation analysis. For this purpose a robust pre-processing scheme based on the wavelet analysis of the heart sounds is introduced. Then, three feature vectors are extracted depending on the cepstral coefficients of different frequency scale representation of the heart sound namely; the mel, bark, and linear scales. Among the investigated feature extraction methods, experimental results show that the mel-scale is the best with 94.4% correct identification rate. Using a hybrid technique combining MFCC and DWT, a new feature vector is extracted improving the system's performance up to 95.12%. Finally, canonical correlation analysis is applied for feature fusion. This improves the performance of the proposed system up to 99.5%. The experimental results show significant improvements in the performance of the proposed system over methods adopting single feature extraction.