Research On High Intelligibility Speech Enhancement Algorithms Based On Cross Terms

Modern speech interaction technology is developing rapidly,while the inevitable background noise is limiting the quality of speech interaction.Speech enhancement techniques are the process of removing as much distracting noise as possible and restoring clean speech.While traditional speech enhancement algorithms have been designed around improving speech quality,recent research has shown that improving speech intelligibility is the key to improving the quality of speech interaction.This paper focuses on the cross-term problem in transform domain speech enhancement algorithms,to obtain more efficient speech intelligibility enhancement solutions through reasonable assumptions and analysis.The main research elements are as follows:A new classifier of noise interference types is designed for the application of a bistatic Wiener filter in the discrete cosine transform(DCT)domain,where noise interference is divided into destructive and constructive noise to avoid the error problem arising from the assumption of zero cross terms,for which two types of filters are used for speech enhancement.However,the type of noise interference needs to be judged before the filtering process is carried out,and the fundamental reason for the different types of noise interference is the similarity and difference in signs between clean speech and noise.Therefore,this paper proposes to discriminate the type of noise interference by identifying the differences and similarities between the symbols of clean speech and noise and to determine which filter to use for effective filtering by comparing the mean square error produced by the two filters.After experimental verification,the classification algorithm proposed in this paper works better than existing probabilistic noise interference type classifiers and can effectively improve the intelligibility of the algorithm’s output speech.In the Discrete Fourier Transform(DFT)domain,existing speech enhancement algorithms mostly assume that the cross term between clean speech and noise is zero,i.e.the background noise is constructive noise and the amplitude of noisy speech is always greater than that of clean speech.In this paper,the gain factor of conventional speech enhancement algorithms is investigated,while the widespread existence of destructive noise in noisy speech is analyzed.Three speech enhancement algorithms with constraints are used to process noisy speech,assuming that the type of noise interference is known,to investigate the improvement in the intelligibility of the algorithm’s output speech considering the presence of destructive noise.Simulation results show that conventional speech enhancement algorithms are inherently deficient in improving speech intelligibility and that incorporating destructive noise into the design of speech enhancement algorithms can significantly improve the intelligibility of speech output from speech enhancement algorithms.Based on the results of the above study,a new noise interference type discrimination method is proposed for the application of DFT domain speech algorithms,which simplifies the calculation to find the probability of occurrence of two noise interference types,destructive noise,and constructive noise,by approximating the probability density function of Nakagami distribution instead of the probability density function of Rician distribution,and obtains highly intelligible speech after combining the constraints The algorithm is then combined with constraints to obtain a highly intelligible speech enhancement algorithm.The effectiveness of the algorithm is verified through simulation experiments.