Short Utterance Speaker Recognition Research Based On Weighted Principal Components

Speaker recognition under short utterance condition is an important research area for speaker recognition technology to be applied in practical application.In a practical interactive system,it is not supposed to require the users to speak too much so as to be user-friendly.Thus,to meet the specific performance requirements,it's important to choose the test samples of appropriate length and enhance the speaker's personality information and suppress the semantic information.After discussing the statistical recognition model and speaker characteristic parameters,this thesis studies the relationship between the recognition performance and the length of test samples,and proposes an exponential function model to describe it in the terms of the dependency between the two.To select an effective length of test samples depending on the function model will contribute to optimizimg system performance.After analyzing the Principal Component Analysis method,this thesis proposes a modified add-and-remove components method to evaluate the contribution rate of every component.In considering of the effect of semantic-separation and contribution rate of the principal components,this thesis proposes a weignted principal components algorithm and it workes to enhance the speaker's personality information and suppress the semantic information.It shows that the algorithm helps to improve the recognition rate when choosing the test samples of appropriate length.By the expriments and analysis,it shows that the relationship between the recognition performance and the length of test samples can be described by an exponential function model,and the good system performance will be achieved when the length of test speech is within 2-6 seconds.The weighted principal components algorithm proposed is able to achieve a 2.88% increase when test speech is within an effective length interval of 2-6 seconds.