Channel Compensation For Speaker Recognition

Automatic Speaker Recognition belongs to the area of pattern recognition and intel-legent computer interface. The ultimate purpose of Speaker Recognition is to makecomputer identify human being. Now many system's performance is fairly good inlaboratory. However they usually perform too much worse to be used in real environ-ment because of disturbance of various noise and unknown factors. Therefore, how toremove or alleviate noise e?ectively and improve performance of Speaker Recognitionsystem becomes an important task.The derivation of channel robustness can come down to the mismatch betweentraining and testing environment when perform speaker recognition. Generally speak-ing, channel compensation methods are various and can be roughly classified intothree categories: feature domain, model domain and score domain. The feature do-main channel compensation focus on the improvement on feature extraction and theprocess on speech feature, which make features vary unconspicuously as environmentchanged. The model domain method adjust models to meet the change of environ-ments.The last domain method use score normalization algorithm to alleviate the dif-ference among scores.This thesis firstly give a brief introduction of speaker recogni-tion task, especially text-independent speaker verification. Then propose the baselinesystem–Gaussian Mixture Model-Universal Background Model in detail.In chapter 3, we propose a channel compensation algorithm naming eigenchan-nel, this algorithm can be implemented in both feature domain and model domain. Itsimplifies the joint factor analysis model, significantly reduced complex and compu-tation to make real-time applications possible. In the NIST 2006 speaker evaluationtask, the model domain eigenchannel system can get a decrease of 48.4% comparedto the baseline system's equal error rate(EER). Features generated by feature domaineigenchannel algorithm can be used for all other speaker recognition systems, it cangreatly enhance the promotion of the algorithm.For the need of NIST 2008 speaker evaluation core test, Chapter 4 of this the-sis realize the joint factor analysis model, As to the limitation of estimating speakerand channel space, we produced a parallel estimation of speaker space model param-eters and serial estimation of channel space model parameters algorithm. Besides theremarkable reduction of complexity and computation, this algorithm can get better per- formance. In the NIST 2006 SRE corpus, the equal error rate(EER) of the proposedsystem can reduce by 69.5% against the baseline GMM-UBM system.It is also the bestsingle system in NIST 2008 speaker recognition evaluation.