Research On Speaker Verification In Complex Background

The speaker verification in complex background refers to the identity confirmationof speaker under complex background. Complex background, contains not only’speakers voice, but also background music，noise and a variety of impurities. Therefore,the work of speaker verification in complex background is divided into two parts,speech signal separation and pure speech signal identity confirmation.Current speech signal separation includes two kinds of methods, Auditory SceneAnalysis (ASA) and Blind Source Separation (BSS). BSS processes mixed signals andextracts the independent component (also known as Independent Component Analysis(ICA)). There are many improved ICA methods and one of them with good separationperformance and fast convergence speed called FastICA has been widely used in thefield of signal separation. This article uses the basic principle of FastICA algoirthm，andintroduces Bayesian estimation based on MCMC. The method improves the separationperformance when the signal is non sparse signals.For speaker veirfication model selection，we adopt an improved GMM model calledglobal background model-Gaussian mixture model (GMM-UBM) as our analysis objectof speaker confirmation has nothing to do with the text. Based on a large number oftrained speech signals, this model gets the higher-order GMM which is used to descirbethe distirbution of characteristics. To improve the robust performance of the model，thearticle uses two channel compensation methods，Feature Mapping and MAP. Bothmethods have achieved good results.Finally^based on the above methods, this paper completes the research work ofspeaker verification system under complex background introduced above using VisualStudio2010and Matlab2014a. At the same time, the paper introduces the algorithmand the steps of system implementation in detail. What is more, this paper also showsthe research results and problems need to be considered and analyzed.