A Study On Noisy Robust Problems In Automatic Speech Recognition By Using Model Compensation

Robust speech recognition for noise environment has become one of the most important research aspect for state-of-the-art speech recognition systems.Although currently the speech recognition system used in laboratory has achieved satisfied system performance, in real application environment,the performance of speech recognition system will decrease a lot because of different kinds of noise distortion.Under this background, in this thesis,the work is expanded according to how to improve system performance in noisy environment by compensating acoustic model parameters to make the acoustic model match noisy speech more accurately.It provides a systematic and in-depth research in this topic,and introduces our innovations in model parameter compensation and training method of acoustic model.Firstly,this thesis proposes a novel acoustic model parameter compensation method in noisy environment-unscented transformation(UT) compensation.To solve the compensation problem for nonlinear transformation of acoustic model parameter in noisy environment,Parallel Model Compensation(PMC) and Vector Taylor Series (VTS) compensation are two mainly used model parameter compensation methods which can only reach accuracy of first order linear approximation.Unscented transformation is used to improve the performance of extended kalman filter in the field of automatic control because it can reach accuracy of second order linear approximation. UT compensation can make us to estimate more accurate noisy acoustic model parameters corresponding to static feature of speech and we also proposes several kinds of engineering implementations to improve the efficiency up to six times faster than usual implementation of UT compensation.The following experiment results show that UT compensation has significant advantage than conventional VTS compensation method.Secondly,we try to expand UT compensation to deal with the acoustic model parameter corresponding to dynamic features of speech.Because of the complicated computation of dynamic features of speech,it is difficult to calculate the compensation of dynamic parameters of acoustic model in noisy environment.In this thesis,we derive the exact nonlinear distortion function of dynamic parameter following precise relationship between dynamic feature and static feature.Then UT compensation is also used to approximate the nonlinear distortion function.Experiment results show that we can get further improvement by using UT compensation for dynamic parameters of acoustic model.Lastly,encouraged by successful progress in speaker adaptive training in speech recognition,we proposes a novel acoustic model training framework by combining the adaptive training and model parameter compensation method.In this thesis,using VTS compensation as an example,obeying the criterion of maximizing the likelihood function of compensated acoustic model to noisy speech database,we train the pseudo clean acoustic model directly.Noisy compensation model adaptive training is a innovational model training method to make the pseudo clean acoustic model to absorb the compensation error of nonlinear transformation approximation.Experiment results show that this method can decrease the recognition error of speech recognition system significantly.And now it has become one of the standard configurations for noisy robust speech recognition system.