Study On Noise Reduction Of Microphone Array Signal In The Reverberation Environment

Due to the background noise and the reflections from walls and other objects,speech signals recorded with a distant microphone in a room usually contain noise and reverberation,which degrade the fidelity and intelligibility of speech,resulting in the decrease of speech recognition rate.Therefore,how to effectively eliminate the noise and reverberation to make the output speech signal as close as possible to the pure speech signal is an urgent problem to be solved.This paper mainly studies the noise reduction and dereverberation algorithms based on microphone array.Two noise reduction and dereverberation algorithms are proposed based on the Habets dual microphone noise reduction and dereverberation algorithm.One of the algorithms is based on the fixed beamforming,and the other is based on the transfer function generalized sidelobe canceller.The study of blind estimation of reverberation time is also focused on in this paper.The main research works are as follows:(1)The characteristics of noise and reverberation,the structure of microphone array and the theory of beamforming are analyzed and studied,to lay the foundation for the subsequent algorithm improvement.The types of noise and reverberation,the related characteristic parameters and statistical models are studied,to understand the characteristics of noise and reverberation.The microphone array is studied,including the array topology,the far and near field model and the time delay estimation.The algorithms in this paper are based on the far-field model.The beamforming algorithms are analyzed,and the advantages and disadvantages of different kinds of beamforming algorithms are summarized.(2)The blind estimation of reverberation time has been studied in this paper and presents a new algorithm,which is based on linear prediction.Traditional reverberation time blind estimation has large amount of calculation and its precision is not high.An improved algorithm based on linear prediction has been proposed,the proposed algorithm utilizes the speech signal directly without any other conditions,and proposes a fast bisection method to solve the maximum likelihood equations.The results show that the new algorithm not only improves the estimation accuracy but also meets the real-time requirement.(3)Based on the research of the Habets dual microphone algorithm,a new algorithm based on fixed beamforming is proposed.The Habets dual microphone algorithm is extended to four microphone algorithm,the four microphones constitute a fixed beamforming,and the post filter using the log spectral amplitude of single channel algorithm.The key technologies such as the estimation of noise and reverberation power spectrum are also studied,and the effectiveness of the algorithm is verified.(4)To solve the problems of the noise reduction and dereverberation algorithm based on fixed beamforming,a new algorithm based on transfer function generalized sidelobe canceller is proposed.Replace the fixed beamforming to the transfer function generalized sidelobe canceller to make the system to realize the adaptive filter.A direct path compensation strategy is applied to prevent the reverberation power spectrum from over estimation.A wiener filter is used to reduce speech distortion of output.The simulation data and real data are used to verify the validity of the algorithms.The recognition of the improved algorithm's output improved significantly.