Research On Robust Speaker Recognition Algorithms In Noisy Environment

There is rich information conveyed in spoken language,including language information(linguistic contents etc),speaker information(identity,emotion,physiological characteristics etc.),environmental information(background,channel etc.)and so on.Speaker recognition refers to recognizing persons from vector sequences extracted by their voices that reflect their identities.In recent years,the performance of speaker recognition is quite high in the laboratory environment.However,they suffer substantial performance degradation if they are operated in an adverse environment including background and channel distortion noises.In order to put the speaker recognition technology into the practical application and reduce the influence of noises on the performance of speaker recognition,this dissertation designs a robust speaker recognition based on MFCC feature-space and VQ model-space,and make an in-depth study on the anti-noise robust speaker recognition algorithms in the real environment.As the existing anti-noise technologies,such as the speech enhancement in the signal-space,feature normalization in the feature-space and model compensation in the model-space,are mostly done by estimating ambient noises based on the SNR,and then de-noising or compensating for specific noise.This method is effective as long as the noise is stationary,such as the assumption of the speaker recognition system on a telephone.However,it may not be effective for the real-environmental situation since the environmental noises usually change dynamically.To solve this problem,two kinds of noise reduction algorithms are studied in the dissertation,which solves the problem of the destroyed signal features and mismatch of training and testing environment caused by noises respectively.Firstly,the robust algorithm based on the speech enhancement and missing feature theory in the feature-space are studied in the dissertation,generating missing feature masks automatically from the enhancement step for judging reliable feature.And in order to enhance the input signal and further provide useful information for the missing feature mask,we use an optimally modified log-spectral amplitude(OM-LSA)speech estimator and a minima controlled recursive averaging(MCRA)noise estimation approach for robust speech enhancement.The MCRA noise estimate is unbiased, computationally efficient,robust in respect to the input SNR and type of underlying additive noise,and characterized by the ability to quickly follow abrupt changes in the noise spectrum.The OM-LSA estimator demonstrates excellent noise suppression,while retaining weak speech components and avoiding the residual noise phenomena.Besides,in order to solve the problem of training and testing mismatched caused by noises,the dissertation studies the Parallel Model Combination(PMC)algorithm in the model-space.The speaker recognition system in the dissertation estimates the additive and convolutive noises from the speech signals,and updates the speaker models(derived from training signals generated in clean conditions)to match the noisy environment.Finally,the dissertation tests the robust algorithms used in speaker recognition system under the white and Babble noise environment,and results follows: 1.we compare the combined algorithm with the OM-LSA algorithm,and the average PESQ value is improved by 0.335 and 0.419,respectively.2.The algorithm based on the speech enhancement and missing feature theory is superior in low SNR and non-stationary noise environment.3.Experiments show that the system can achieve 100% recognition rate in the case of 30 d B input SNR,in which the PMC algorithm contributes 10.9% accuracy to the system recognition rate.