Research Of Speaker Verification On Feature Compensation

With the social progress and the development of computer technology, speaker recognition is more widely used as an identity authentication technology due to its unique advantages. In the field of speaker recognition, text-independent speaker verification technology is currently regarded as the research focus and difficult point. Especially in the recent years. The background noise, complex transmission channel and diverse types of microphones led to a bad result which makes a speaker verification system performance more difficult to be further improved.This paper researches the methods of feature compensation on speaker verification by reducing channel mismatch and using secondary feature parameters. We proposed several new approaches, given experimental proof. The main contents are shown as follows:1. We discuss the cepstrum analysis and the extraction process of short-term vocal tract parameters MFCC.The composition principle of gaussian mixed model, MAP adaptive algorithm to get a speaker model, the initial calculation, EM training methods, the universal background model UBM are deeply explored. Speaker verification system is constructed and tested based on GMM-UBM with the given speech database. The impact of MAP adaptive algorithm to adjust different parameters were researched by experiments to see the system performance.2. Unsupervised feature mapping method for channel mismatch compensation is deeply explored. By using principal component analysis to analyze male and female voice, We know the distance of the opposite gender on the basis of the GMM supervector space is very far.Based on the result we proposed a new feature mapping method based on blind channel clustering, to solve the problem of unsupervised feature mapping without adequate corpus cases. Channel mismatch compensation is achieved by using this method on the GMM supervector space.Experiments show that this method overcomes the disadvantages of the poor corpus, and the relative performance has been improved compared with the main system.3. The extraction of pitch is deeply studied, using a band-pass filter to reduce the effects of the majority of the formant. The improved autocorrelation function in software "Praat" is also explored. We proposed a new score fusion method using a threshold on the basis of the traditional linear score fusion. By statistical analysis of test scores on correct and wrong test cases, the feasibility of this method can be demonstrated, the range of threshold can also be estimated. Experiments show that this new system fusion solution is able to protect the good score distribution, making the pitch auxiliary system compensate for the performance of the primary system better.