Speech Enhancement Algorithm Based On Scene Matching For Binaural Hearing Aids

At present,the speech enhancement algorithm in binaural digital hearing aids is mainly based on adaptive beamforming algorithm.This algorithm strongly depends on the environment,and it is not effective for non correlated noise.To solve this problem,a binaural speech enhancement algorithm based on sound source localization and scene matching is proposed.This algorithm is based on the key technologies of speech enhancement,sound source localization and scene matching.The spatial information is extracted by the sound source localization algorithm to ensure the integrity of the spatial information.The noise in different directions from the speaker is removed and the dual channel speech signal is reduced to single channel speech signal.Then the scene matching is used to detect the background environment.The scene matching algorithm is more complex and less suitable at present.So,in this thesis,a scene matching algorithm based on multi-feature is presented.The feature parameters are combined by this algorithm in speech segments and noise segments respectively.The improved weighted minimum distance classifier is used in this algorithm to identify noise type.Then according to the matching result,the corresponding speech enhancement algorithm is used to denoise.In order to achieve the purpose of speech enhancement,the spatial information is restored by the HRTF(Head-related transfer function)module.Speech enhancement algorithms are generally based on the assumption of the characteristics of speech and noise.Therefore,the speech enhancement algorithm has different performance in different noisy scenes.In this thesis,the multi channel enhancement in binaural hearing aids is converted to single channel,which reduces the computational complexity.And the characteristics of spatial information is also retained.This algorithm can deal with different background noise.The research of this thesis has important practical significance to the development and wide application of binaural hearing aids.The main work of this thesis is embodied as follows:(1)Research of the sound source localization algorithm.The existing sound source localization algorithm for binaural digital hearing aids has a good positioning ability,but it can not be applied to the multi scene which needs high accuracy and low complexity of the algorithm in digital hearing aids.A new sound source localization algorithm is proposed based on the analysis of the previous algorithms.Firstly,the binaural sound source signal is decomposed into several channels by Gammatone filter,and the high energy channels are extracted by adding weighting coefficients.Then,the noise in different directions from speaker is removed by calculating the relevant parameters of HRTF and firstly introducing the first GMM(Gaussian mixture model).The computational complexity of the speech enhancement module is greatly reduced through this step.(2)Research of the scene matching algorithm.The recognition precision of the existing auditory scene recognition algorithms is relatively satisfactory,but they can only be applied to several noise scenarios,and it can't meet the performance requirements of digital hearing aids in complex environment.Based on the existing research of auditory scene matching algorithm,a new algorithm for auditory scene matching is proposed.In this algorithm,the speech endpoint detection algorithm based on the band-partitioning spectral entropy and spectral energy is used to divide the noisy speech into speech segment and noise segment.Then the characteristics such as Critical Band Ratio and band-partitioning spectral entropy as well as adaptive short-time zero crossing rate of each segment are extracted for the weighted minimum distance classifier to recognize the noise scenario.(3)Selection of the optimum speech enhancement algorithm for different noises.The noise in the same direction as the speaker is removed by the speech enhancement module.Different noises have different characteristics.The denoising performance of the speech enhancement module can be improved by denoising their respective characteristics.The optional algorithm can be summarized by a large number of denoise experiments.Experiment results show that The proposed binaural sound source localization algorithm based on HRTF and GMM classifier with Gammatone filter decomposition has strong robustness,high accuracy and lower complexity.the proposed scene recognition algorithm based on multi-feature and weighted minimum distance classifier for digital hearing aids has strong robustness and high accuracy.It's suitable to be applied in digital hearing aids.The binaural speech enhancement system based on scene matching has better denoising performance than the reference algorithm.