The Research Of In-Car Speech Enhancement Algorithm Based On Blind Source Separation

Speech plays an important role in people's daily life because of its convenience,speed,efficiency in communication.With the continuous advancement of social technology and the rapid development of artificial intelligence,speech has also gradually become a highly active application in human-computer interaction,compared with the traditional human-computer interaction methods,speech is of more convenience,efficiency and security.So speech interaction has been widely used in various fields such as industrial control,medical assistance,security protection,smart home and so on.However,in the real application scenarios,the speech signal would be inevitably contaminated with enviromental noise,which affects the speech quality as well as fails to complete the normal human-computer interaction.Therefore,speech enhancement,as an effective method to suppress noise and improve the speech quality,are of great research significance and application value.For the specific application scenario of in-car environment,the noise signal is with the characteristics of low frequency distribution,deficiency of prior knowledge,and the mechanism it mixing with speech signals is very complex,therefore many speech enhancement algorithms do not produce acceptable performance in the in-car environment.Based on the analysis of in-car noise and acoustic scene,the convolutive mixture model was employed to describe the mixing process of noise and speech,Aiming to improve the quality and intelligibility of in-car noisy speech,the effectiveness and feasibility of Blind Source Separation(BSS)based speech enhancement was researched.This dissertation is organized as follows:(1)Analysis and modeling of in-car acoustic scene and research on noise estimation algorithm.According to the inherent characteristics of the in-car environment,the source of in-car noise and the in-car propagation path of driver's speech is analyzed to establish the in-car convolutive mixing model.Since most of the speech enhancement algorithms require estimated noise as a priori knowledge for denoising,it is crucial to precisely estimate noise because it's closely related to the performance of the speech enhancement algorithm.The dissertation researched some common noise estimation algorithms on the basis of summarizing some commonly used speech processing theories,including speech endpoint detection algorithm and minimum control recursive average algorithm.(2)Research on speech quality evaluation and speech enhancement algorithms.This dissertation summarized some widely used subjective and objective evaluation principles of speech signal quality,and analyzed the advantages and disadvantages of these evaluation standards.Meanwhile,aiming at the lack of reference source in the objective evaluation standards in the real environment,a small vocabulary speech recognition engine based on Hidden Markov Model(HMM)was constructed in this dissertation,and the speech recognition rate was incorporated into the evaluation system of speech quality with no reference source.As to the research on speech enhancement algorithms,firstly,two classic speech enhancement algorithms,spectral subtraction and wiener filtering,are experimentally analyzed,and their de-noising results for in-car noisy speech were given in this dissertation.Secondly,an improved speech enhancement algorithm based on wavelet threshold function was proposed to overcome the shortcomings of some traditional speech enhancement algorithms.The proposed algorithm can effectively suppress wideband noise as well as improve speech quality.In the end,the dissertation demonstrated the basic theory framework and implementation principle of Independent Component Analysis(ICA),and focused on researching the process of frequency domain blind deconvolution based on complex ICA with negative entropy.The ICA's speech enhancement process can not only better fit the convolutive mixture model,but also make up the deficiency of the existing speech enhancement algorithms in the car environment.(3)Research on in-car speech enhancement algorithm based on convolutive ICA.According to the convolutive mixing of speech signal and in-car noise signals and their non-Gaussian distribution in the frequency domain,the dissertation proposed and optimized a speech enhancement algorithm based on convolutive ICA by maximizing the negative entropy.This dissertation constructed an in-car noisy speech signal corpus in three kinds of acoustic scenes:simulation environment,indoor environment,and real in-car environment,and employed convolutive ICA based on negative entropy to denoise.The experimental results showed that the recognition rate of speech signals enhanced by convolutive ICA are respectively 18.33%,30%and 27.5%higher than that of in-car noisy speech signal,which revealed the effectiveness and robustness of convolutive ICA in car acoustic scene.In the end,the problem of the frame length and frame shift selection which affect the enhancement effect of convolutive ICA was experimentally researched and discussed.(4)Research and implementation of speech enhancement system in complex environment.Based on noise estimation algorithm and speech enhancement algorithms that this dissertation researched,a speech enhancement system with several algorithms and control logics in it is developed using C++ on Windows platform.The system has the functions of speech waveform display,spectrum display,selective speech enhancement,speech playing and so on.The test results indicated that the system not only has good speech enhancement performance,but also obtains strong reliability and robustness.