A speaker recognition algorithm based on factor analysis

Abstract

Channel interference factor for the identification result is prevalent among the existing speaker recognition algorithms. In order to improve the accuracy of the algorithm, the paper utilizes the technique of latent factor analysis(LFA) to deal with the channel factors in the speaker's Gaussian Mixture Model(GMM). In the endpoint detection phase of speaker recognition, the algorithm introduces the GMM for speech modeling to accurately determine the beginning and ending points of the speech segment, and then establish speaker GMM. The algorithm use factor analysis technique to fit the differences between the speaker characteristics space and the channel space, and removes channel factor in speaker's GMM. And then the algorithm extracts GMM super-vectors as the input of Support Vector Machine(SVM) to obtain recognition results. Experimental results show that the combination of factor analysis and SVM can obtain better recognition rate and ensure the robustness of the recognition algorithm. (Abstract)

Automatic Speaker Recognition System Based on Gaussian Mixture Models, Cepstral Analysis, and Genetic Selection of Distinctive Features

“…This subsection presents the results of the effectiveness of the ASR System, and compares them with the results obtained by other authors [ 4 , 5 , 6 , 7 ]. Unfortunately, very often, the implementation of such a comparison is an extremely difficult task, due to the different datasets of voices used by different authors.…”

“…Moreover, the datasets should be tested in the same way so that a comparison of ASR Systems can be reliable, which often turns out to be difficult due to the different methods of testing by many authors of articles. For example, in the light of research conducted on the basis of the TIMIT [ 4 , 5 , 6 , 7 ] dataset, the results achieved by the authors are at least satisfactory, as they allow for the achievement of a higher IR level and a lowering of the EER.…”

“…The authors of subsequent papers [ 4 , 5 , 6 , 7 ] used the TIMIT database [ 8 ] for their research. The results of the studies conducted using the TIMIT database are the most widely available and thoroughly described, and so a comparison to other ASR systems using this database is described in detail in Section 4.5 .…”

“…The results of the studies conducted using the TIMIT database are the most widely available and thoroughly described, and so a comparison to other ASR systems using this database is described in detail in Section 4.5 . In the present work, among others, the following were used for distinguishing features: MFCC [ 4 , 5 , 6 , 7 ] and power normalization cepstral coefficients (PNCC) [ 7 ]. In contrast, speaker classification in the present ASR System was based on GMM-UBM [ 4 ], latent factor analysis with Support Vector Machine (LFA-SVM) [ 5 ], Linear Discriminant Analysis (LDA) [ 6 ], Gaussian probabilistic LDA (GPLDA) [ 6 ], and extreme learning machine (ELM) [ 7 ].…”

“…For the NIST 2002 SRE dataset, EER was reduced by 14.57%; for the TIMIT dataset—by 0.14% (and by 10.19% for speech synthesized with the G.729 codec); and for the VCTK dataset—by 4.18% as compared to the competitive studies [ 2 , 6 , 9 ], respectively. In addition, for the TIMIT base, increases in identification rates (IRs) of 19% [ 4 ], 17.16% [ 5 ], and 0.87% (and 8.46% for speech noises using additive white Gaussian noise (AWGN)) were obtained [ 7 ], shown in Section 4.5 . No results of the current research using the SNUV [ 12 ] dataset were found among the publications tackling the issue of speaker recognition.…”