Model-based Speech Enhancement With Microphone Array

The speech signal is always affected by different noises in the actual communication,leading to a decline in quality.The purpose of the speech enhancement is to extract the pure original speech as much as possible from the noisy speech.Compared to single-channel speech enhancement,which only time-frequency information can be used,the microphone array speech enhancement can utilize the spatial correlation between multiple received signals to enhance useful speech signals in a specified direction and suppress noise or interference in other directions,which is a good solution to the enhance the quality of speech and reduce speech distortion simultaneously.Microphone array speech enhancement methods include fixed beamforming and adaptive beamforming.A common method of adaptive beamforming for microphone arrays is generalized sidelobe canceller(GSC),which has a strong ability to suppress spatially coherent noise and a weak ability to suppress spatial incoherent noise.When non-coherent noise and coherent noise coexist,output of enhanced speech residuals a lot of noise.In addition,the GSC is designed for general input signals without prior knowledge instead of optimizing according to the characteristics of the speech signal.In view of the above insufficiency,this paper proposes a model-based microphone array speech enhancement algorithm,establishes a noise model after correcting the incoherent noise of the auxiliary branch,and combines the prior knowledge of the clean speech model to construct an optimal speech filter filtering the GSC main branch signal.Experiment shows this algorithm brings significantly improved performance than the traditional GSC.The work of this article mainly includes the following aspects:1.Summarize the development history and current research status of microphone speech signal processing models and microphone array speech enhancement algorithms;2.Introduce the theory of GSC and subband beamforming,and speech enhancement algorithm based on statistical model;3.Proposed an adaptive beamforming microphone speech enhancement algorithm based on statistical model.This algorithm uses a clean speech library to train the clean speech cepstrum domain and spectrum domain model firstly,then re-estimate the incoherent noise of the GSC auxiliary branch to establish a noise model.Then,construct an optimal filter to enhance the main branch of the GSC,after calculating the noisy speech model and the filter weights using the prior knowledge of the statistical model.4.The simulation of the proposed model-based GSC algorithm was performed.The best classification number of the model was measured and compared with TF-GSC,Frost,model-based and GSC algorithm based on post-filtering.Experiment shows that under Gaussian white noise conditions,the mean PESQ score of the proposed method is 0.67 points higher than that of the traditional GSC,an increase of 31.4%.Under the environment with Gaussian white noise and disturbing music,the average PESQ score of the proposed method is 0.57 points higher than the traditional GSC,with an average increase of 27.8%.This experiment shows that the proposed model-based GSC algorithm in this paper has better speech enhancement performance in the noisy environment,and the speech quality has been significantly improved.