The Study Of Speech Enhancement Technology For Robust Speech Recognition System

In recent years,with the improvement of recognition rate of automatic speech recognition system,the speech recognition technology previously studied in laboratory has gradually been applied to our daily lives.However,the robustness of automatic speech recognition system in noisy environment has been a hot research,and it also has come a stumbling block in the commercialization of speech recognition schemes.In order to improve the robustness of automatic speech recognition system in noisy environment,this thesis studies the speech enhancement technology with an aim of better combating the adverse noise.Based on the performance and complexity of the speech enhancement algorithm,this thesis takes the speech enhancement algorithm based on statistical model mainly studies.Two works of this thesis are summanized as follows:Firstly,different from the traditional Gaussian distribution hypothesis,this thesis assumes that the speech discrete Fourier transform coefficients are better modelled by super-Gaussian distribution(Chi distribution),and a generalized weighted beta-order spectral magnitude estimator with Chi priori was derived.We can find that the generalized weighted beta-order spectral amplitude estimator is a generalized estimator.In order to get more noise suppression and less speech distortion,we obtain a set of adaptive parameters in the improved Bayesian estimator.In order to get the adaptive parameter Beta,we propose an novel voice activity detection method in noisy environment.Combining the VAD(voice activity detection)algorithm with the prior SNR(signal-to-noise ratio),a dynamic parameter Beta is obtained.To obtain the adaptive parameter P,we introduce a sigmoid function.The resultant adaptive Bayesian estimator can obtain different noise suppression for different speech signals.The simulation results show that the proposed algorithm is more effective than the other compared algorithms,and can get higher segment SNR and higher recognition rate in automatic speech recognition system.Next,with the speech presence uncertainty(SPU)estimation,a novel Bayesian estimator incorporating the SPU with Chi priori was derived.Moreover,based on the traditional decision-directed method and temporal cepstrum smoothing method,we proposed a hybrid prior SNR estimation algorithm.The simulation results show that the proposed algorithm has better robustness compared with the contrast algorithm in noisy environment.