Speech Enhancement Of Dual Channel DNN Hearing Aid Based On Human Hearing

Language is a bridge of human communication,but for the deaf,hearing loss has a serious impact on their daily communication.According to the survey,nearly 400 million people with hearing disabilities around the world.Digital hearing aid is an effective measure to solve the problem of hearing impairment,and in recent years,Deep Neural Network(DNN)has been successfully applied in hearing aid with its advantage of expressing complex nonlinear relationship between speech.Therefore,this paper improves the speech enhancement algorithm in digital hearing aids through the optimized DNN to further improve the reconstructed speech quality.The main work of this paper is as follows:(1)This paper expounds the research status and significance of digital hearing aids,and briefly introduces the working principle and key technology of digital hearing aids.This paper focuses on the speech enhancement part,and describes the existing research results in detail from the two speech enhancement algorithms based on DNN and beamforming.(2)Aiming at the problem of poor speech quality caused by the feature selection of neural network speech enhancement algorithm can not fully represent the nonlinear structure of speech,a dynamic feature combined with a adaptive ratio mask optimization neural network speech enhancement algorithm is proposed.Firstly,three features of noisy speech are extracted and stitched to obtain static features.Then,the first and second differential derivatives are obtained to capture the instantaneous signal of speech and fuse it into dynamic features.The combination of dynamic and static features completes the internal complementary of features and reduces speech distortion.Secondly,in order to achieve the best intelligibility and clarity of enhanced speech at the same time,a new adaptive mask is proposed,which can adaptively adjust the energy ratio of speech and noise,as well as the ratio of traditional mask and square root mask;and the mask value in each channel is modified with gammatone channel weight to imitate human auditory system,so as to further improve the intelligibility of speech understanding.Finally,the simulation results show that the SNR,speech quality and short-term objective intelligibility of the proposed algorithm are higher than those of other algorithms in the literature,which verifies the effectiveness of the proposed algorithm.(3)In order to improve the auditory perception of reconstructed speech in the speech enhancement module of digital hearing aid,a two-channel DNN time-frequency masking speech enhancement algorithm based on the advantage of single and multi-channel signal processing is proposed in this paper,which is based on the Bark domain loss function optimization.Firstly,in order to improve the problem that the traditional time-domain optimization of DNN loss function does not accord with the auditory perception of human ears,a weighted loss function of Bark domain is proposed in this paper to improve the objective quality of enhanced speech by optimizing DNN in the psychoacoustic domain.Secondly,in order to improve the ability of the algorithm to suppress directional noise,the direction vector localization method based on adaptive mask is proposed in this paper.By accurately calculating the spatial covariance matrix and direction vector of speech and noise,it can accurately locate the target sound source in the environment with noise and reverberation.Finally,the experimental results show that the reconstructed speech has better speech quality and intelligibility compared with other single and multi-channel speech enhancement algorithms.