Speaker Rcognition Using Wavelet Packet Decomposition Based On MPEG-I

Speaker recognition is the task that is used to identify or verify who is speaking by analyzing and recognizing specific information extracted from voice signals. As an important branch research field of speaker recognition, text-independent speaker recognition has carried out all over the world and plays a more and more important role because of its ease-to-use and highly popular application in the information technology and it will have a significant future. Because of eliminating the effect of reliant of text, the key of research is focusing on features extraction. And finding new speech feature and combination of existing speech features are hotspots of research on speech feature extraction.In this paper, the focus is to find out a method of feature extraction which has a high recognition rate and robust. The main work has three aspects:(1)Research on the component of voice signal based on the apperceive characteristics of ear. And we decompose signals according to MPEG-I psychoacoustic model, and eliminate the interference from those signals under global masking threshold. The experiments confirm that this approach has effectively improves the system's recognition rate and enhances the robustness of the identification.(2) Using the characteristics of multi-resolution of wavelet packet we carry multi-scale decomposition on voice signal. This method replaces the Mel-Spaced filter and FFT in the process of feature extraction of MFCC. It's proved that this method simplifies the process of feature extraction and WPTC perform better than MFCC in speaker recognition.(3)According to the characteristics of RBF, we design a special structure of neural network: design an independent sub-net for each speaker. The method decrease the interference of feature vector of others speakers.