Reaserch Of Speaker Verification On Channel Compeansation

Text-independent speaker verification is one of the main areas of speaker recognition. In recent years, with the increasing computing power of computers and portable devices, as well as the rapid development of speaker recognition technology, speaker recognition shows great research value, broad application prospects. Therefore, the research on speaker verification technology, is of great practical significance.The state of art speaker verification systems are mostly based on probabilistic model. Probability statistic models are able to describe acoustic characteristics of speakers in the feature space, which can contribute to very good results in speaker verification. However, due to the influence of background noise, channel mismatch and other issues, performance of speaker verification system is difficult to make further improvement. This paper main focus on model domain compensation methods for speaker verification, and discuss the ways to eliminate the channel mismatch, exploring the methods to get discriminative training. Speaker verification technology based on GMM-SVM, Total Variability space and G-PLDA are discussed respectively. This article mainly focuses on the content:First, GFCC commonly used in Computational Auditory Scene Analysis(CASA) is applied to speaker verification. And some improvements are made, such as replacing logarithm with the square root of10when feature compression, using26channel Gammatone filters when extracting feature instead of64channels. Experiments show that GFCC with26channel Gammatone filters can achieve better performance. What’s more, GFCC features based on the square root of10in feature extraction outperform PLP, MFCC features.Next, GMM-SVM speaker verification system has been constructed, with PCA transformation in acoustic parameters ahead of GMM modeling and SVM discriminative training. And also, proposed a new method to make full use of GMM-UBM parameter set as SVM input. Experiments show that this method can boost the system performance obviously against the origin system.Finally, regularizations and normalizations, based on I-vector, including Whitening, Length normalization, Linear Discriminant Analysis (LDA) and Within Class Covariance Normalisation (WCCN), are exploited to eliminate the channel effects and background noises. Channel compensation technology’s impact on the experimental results is deeply analysed, and proposed that LDA or G-PLDA transformation after Whitening and Length normalization, can make the performance of I-vector system much more better.