Research On Speaker Recognition Based On MFCC And PSO-BP Neural Network

With the development of digital information, speaker recognition is a kind of biometric authentication. Compared with the technology based on traditional linear system theory, speaker recognition based on the neural network has a strong ability of pattern classification and robustness to the incomplete information, which becomes a hot spot in the area of speaker recognition. Artificial neural network has the advantage of autonomous learning in fuzzy information, according to imitating animal behavior characteristics.This dissertation firstly introduces the development process and research status of speaker recognition, artificial neural network and the optimization algorithm. The preprocessing of speech signal is described in details, including wavelet denoising, pre-emphasis, traditional dual-threshold endpoint detection. Filtering the noise and the silence could offer the useful speech signals which can be used in the features extraction. Compared with the MFCC feature parameters, the one based on spectral subtraction speech enhancement has good noise robustness.For the traditional speaker recognition networks which need large amounts of training data, large storage and poor robustness, this paper presents the back-propagation algorithm. The BP neural network which has the advantage of improving the system performance with the experience, adjusting the parameters with self organization and self adaptive learning. In order to solve the question that BP Neural Network has slow convergence speed and is apt to fall into the minimum value, the particle swarm optimization algorithm is adopted in the network. The weight value and the threshold value of the BP network could be trained by using the PSO. Therefore, the local convergence can be effectively prevented and the global search can be speeded up. Compared with the traditional speaker recognition networks, the performance of speaker recognition system based on PSO-BP neural network could be greatly improved, the recognition rate and training speed of the system based on PSO-BP network are remarkably increased verified by experiment simulation.