Research on Wavelet Analysis Theory Applying in Speaker Recognition
|Course||Measuring Technology and Instruments|
|Keywords||Speaker recognition Wavelet analysis Speech improvement Bark-scaled wavelet package transform Feature extraction|
Speaker recognition is one of the main applications of speech processing techniques and a hot issue in the field of biometrics as well as. The solutions to the existing problems in the speaker recognition systems depend basically on the effect of analysis of speech signal. Wavelet analysis is a new time-frequency localization method with many advantages that can not be found in other methods, being able to describe speech signal more subtly and to capture the unstable information in the speech signal. Therefore, this paper mainly studies the application of wavelet analysis theory in speaker recognition.At the beginning of the paper, the theory of wavelet analysis is introduced, especially the multi-analysis of wavelet transform and the partition technique of frequency band of wavelet package transform. Then, speaker recognition model (GMM model) and experimental data are introduced. As experimental platform, a speaker identification system is presented, which is finished by MATLAB and VC++ programming language.The study of wavelet analysis theory in speaker recognition in this paper is mainly about the following two aspects: the application of wavelet-based speech improvement in speaker recognition and the design of feature parameters based on wavelet package transform. The details are as follows:(1) The principle of wavelet-based speech improvement is expatiated and a new arithmetic is presented. Experiments show that the new arithmetic not only has excellent effects on speech improvement but also has potential to improve robustness of a speaker recognition system in noisy environments.(2) The conventional feature parameters applied in speaker recognition are summarized and those disadvantages are analyzed. Then, a method named Bark-scaled wavelet packet transform of speech signal is proposed in this paper. The method uses wavelet packet transform to simulate the Bark field frequency perceptual properties of human auditory system and a new structure of wavelet packet decomposition is devised according to the frequency <WP=6>distributional characteristics of speaker information. On the basis of the above, a new feature for speaker recognition is presented. The ability of the new feature to capture speaker information conveyed in the speech signal is compared to the widely used MFCC. Experiment shows that the new feature is superior to MFCC in robustness and it has potential in improving the performance of speaker recognition system.