Research On Robust Speech Recognition In Noise Environment

In recent years, with the development of science and technology, all kinds of intelligent products are going into people’s life, such as smart phones, intelligent home products, etc. Along with the rise of the intelligence market, man-machine communication requirements are also getting higher and higher. These all promote the development of speech recognition technology, and make it become an important research direction in natural language processing field. At present, speech recognition technology can get highly accurate in the recognition of laboratory voice, but once mixed with the noise, it is difficult to match the test speech, making the system performance deterioration in real life environment.The goal of robust speech recognition is to maintain the stability of the system in different environments. The commonly used nonlinear compensation models are VTS(Vector Taylor Series), DPMC(Data Driven Parallel Model Combination) and UT(Unscented Transformation), etc. In this thesis, the robust speech recognition technology is discussed from two aspects: the feature compensation and the model space compensation. A speech enhancement algorithm and VTS algorithm(based on EM noise feature compensation) are studied in this paper. In this compensated VTS algorithm, if there is noise in the speech after feature compensation, these noises will pass back and accumulate in the system, and the parameters of the model will be relatively easy to implement, so the VTS compensation based on model space and GN(Gauss-Newton) algorithm are also studied in this paper.Experimental data show that the performance of model space compensation is significantly better than the feature compensation. GN algorithm can estimate more accurate noise parameter and get higher convergence rate. In the training process of model parameters, the iterative times of GN algorithm is less than the VTS compensation algorithm(based on EM noise estimate). The GN algorithm can effectively save the calculation of the system, while the average decline of 0.53%-2.34% than the previous system error rate.