Research On Speaker Recognition Algorithm Based On Cepstrum Feature

Language is the most basic way of human information exchange,and with the development of information technology,human-computer interaction has become a new and urgent demand in science and technology promotion.Speakers identification technology with fast,effective and low cost advantages,has become widely accepted one of the important biometric authentic technology.The technology is widely used in the information authentication,judicial investigation,e-commerce,document security and other aspects.Speaker recognition,also known as voiceprint recognition,is one of the important areas of speech signal processing technology.Although the speaker recognition technology has achieved some results in theoretical research,the actual environment with high noise and high distortion can cause the recognition performance of the speaker recognition system to drop sharply.Speech feature parameter extraction is the most critical part of the speaker recognition algorithm.Therefore,the current research hotspot is how to extract the recognition performance and distinguish the superior performance characteristics in the high noise environment.Based on the existing speaker recognition algorithm,this paper makes an in-depth study on the feature extraction and weighting,filter design and other aspects of signal feature extraction,and puts forward their own solutions.1.In view of the high variance and latency of the traditional Mel spectrum,the Mel filter group simulates the poor performance of the human ear,the poor resistance of the MFCC,and the static characteristics of the Mel spectrum characteristic parameters.Multi-window estimation and Gammatone filter group,the speaker 's personality feature information is preserved to the maximum extent while reducing the variance and noise reduction of the spectrum,so as to improve the recognition performance of the speaker recognition algorithm.2.For the traditional Mel filter group,the distribution of low frequency and high frequency sparse distribution is not consistent with the spectrum distribution of the effective information of the normal sound,and the number of fixed filters is not suitable for changing the speech signal.The design is based on the Fisher principle and the Gammatone filter.Adapting to the improved filter bank,and propose a speech recognition algorithm based on the adaptive improved filter bank.