Speaker Recognition Based On Swarm Intelligence And Blind Source Separation

Speaker recognition has got excellent recognition rate in the clean voice signal, but when effected by environmental noise, the clean speech signal will be distorted, this made the training environment didn't match with the testing environment, and recognition rate of system is influenced seriously. It is a difficult point how to extract voice signal parameters which indicate individual feature of speaker and design effective classifier, which prevent speaker recognition system from applying into real environment.Based on the above drawback, stating from the two aspects of voice feature extraction and design of classifier, the paper propose the solution and demonstrate they feasibility by experiments.1. The recognition rate of speaker recognition will decline dramatically under the mixing noise environment; we apply the improved Independent Component Analysis (ICA) into de-noising speech signal. The traditional searching scheme of ICA is always gradient-based algorithm; however the convergence and the performance of it are depend on the choice of learning step size. To overcome the drawbacks, an efficient improved ICA algorithm which is based on particle swarm optimization (PSO) is presented in the paper. And the Mel-frequency cepstral coefficients (MFCC) is enhanced by the improved algorithm, simulation results show the method is effective for users to get optimal resolution to the mixed noise voice signal.2. Aiming at the shortage of Support Vector Machine (SVM) slow practice speed in the case of large sample, this paper introduces weighted optimal position strategy to improve Quantum Particle Swarm Optimization (QPSO) algorithm, processes coding for voice parameters by improving Michigan coding scheme, and constructs new classified rule fitness function to realize designing of classifier based on weighted quantum particle swarm (WQPS-classifier). Application results of speaker recognition show that this classifier has better performance of noise immunity and recognition speed.