Research Of Speaker Recognition Algorithm

Speaker identification is one of the biometric authentication technologies,which uses the information contained in a person's speech signal for feature extraction,and also uses the features to identify or confirm the speaker.With rapid development of the Internet and information technology,speaker identification technology will gradually become a focus in research field.Among the numerous speaker identification methods,the improved method of traditional MFCC parameters and Gaussian Mixture model in speaker identification system is studied in this paper.The main content of research is as follows:1.The methods of feature extraction and model building technology for traditional speaker identification systems are systematically introduced.For feature extraction,we focus on the study of MFCC and LPCC parameters.The modeling part lists Dynamic Time Warping,Hidden Markov Model,Gaussian Mixture Model and Artificial Neural Network.2.On the basis of traditional MFCC feature parameters,some new algorithms are proposed in this paper.In order to shorten the speech and reduce the large amount of calculation,endpoint detection algorithm is used and it is able to remove the silence part to implement the requirements.On this basis,the triangle filter is replaced by Gaussian filter group to effectively provide smooth transition between adjacent cabin and improve identification accuracy.3.In the use of Gaussian filter group to improve traditional MFCC algorithm,pitch detection is added to dynamically change the control of variance in Gaussian filter group.Transforming the extracted pitch frequency to Mel frequency and use it to dominate variance control of Gaussian shaped filters.Constructed better representation of different speaker period to character the vocal cord vibration,then the implementation of Dynamic MFCC process is achieved.Experimental results indicate that proposed parameters can significantly improve recognition rate.In the speaker identification system,the proposed method has great value in speaker identification.4.Gaussian mixture model is used for speaker identification in this whole paper.One of the advantages of Gaussian mixture model is that it is able to take advantage of multiple Gaussian distribution and use it to fit different speaker feature vector space.And GMM is the most commonly used model in speaker identification system.Experiments show that both algorithm improve recognition rate.Indicated the methods are effective.