Research And Implementation Of Speech Enhancement Algorithm Based On Beamforming And Dictionary Learning

Speech is an information carrier of human communication and an important way of human-computer interaction.However,in real life,the ubiquitous noise will cause interference and pollution to the voice signal,which seriously affects the quality of voice communication and even causes unpredictable errors.Speech Enhancement(SE)refers to a technology that after the speech signal is interfered by noise,the original speech signal is extracted from the interfered speech as far as possible,and the noise interference is suppressed and reduced.Speech enhancement by the number of microphones can be divided into two types: single-channel speech enhancement and multi-channel microphone array speech enhancement.The microphone array can utilize spatial information and have better suppression effect on noise in all directions.However,the disadvantage is that it requires multiple microphones,and Single channel speech enhancement has this advantage over microphone arrays.Therefore,No matter which kind of speech enhancement we study,it is of great theoretical and practical value to study single channel speech enhancement or multi-channel speech enhancement.Embedded systems have developed rapidly in recent years,it is possible to implement some complex speech enhancement algorithms.This paper mainly studies the microphone array beamforming speech enhancement algorithm and its implementation on DSP,and Joint dictionary learning in single channel speech enhancement is also studied.This article mainly does the following research content,and has the following innovations:(1)In this paper,DSP implementation of microphone array speech acquisition and processing system is studied.Use Texas Instruments The 66AK2 G Evaluation Module(EVMK2G)and K2 G Audio Daughtercard(AUDK2G)developed to build a set of array voice acquisition and processing system.A set of driver program is developed based on C programming language to realize real-time acquisition and playback of speech signals.This set of drivers includes: power-on processing program and system initialization program.On the premise of ensuring the system functionality,the software optimization is completed to reducing the code size of our C programs and improving the speed of the system.(2)This paper studies the beamforming speech enhancement method through several sets of microphones in the case of multi-channel and implements the algorithm in hardware.Aiming at the weakness of beamforming algorithm in the low frequency part of noise suppression,in this paper,a method combining beamforming algorithm and spectral noise reduction algorithm is proposed.In the middle and high frequency part,the beamforming algorithm is used to suppress the noise,and in the low frequency part,the spectral noise reduction algorithm is used to suppress the noise,In this way,high frequency noise and low frequency noise can be filtered out together.We use MATLAB software to simulate and verify the mentioned methods,and the method is deployed on the microphone array speech acquisition and processing system,and a set of four channel microphone array speech enhancement embedded system is realized.(3)A single channel speech enhancement algorithm based on joint dictionary learning is studied.In order to solve the problem of source confusion in the single channel speech enhancement algorithm of joint dictionary learning,this paper proposes a new objective function for dictionary learning.By adding the constraint items projected on the noise sub-dictionary of speech signal and the correlation between dictionary atoms,the trained dictionary is more distinguishable,and the problem of source confusion is reduced to a certain extent.Through experimental simulation,a group of optimal algorithm parameters are obtained under four different noise scenarios.By comparing the experiment with the traditional algorithm and the joint dictionary learning algorithm,the proposed algorithm has better speech enhancement effect.