Drictional Speech Enhancement Algorithm Based On Deep Learning For Binaural Hearing AIDS

In order to improve the performance of digital hearing aids,the comfort and intelligibility of wearing hearing aids for patients with ear injuries is improved.This paper proposes a directional speech enhancement algorithm based on deep learning in digital hearing aids,the algorithm starts from the key technologies of sound source localization,beamforming and speech enhancement in binaural digital hearing aids.Firstly,the head-related transfer function is used to extract the binaural spatial feature information,and the spatial information of the sound source is determined by two-level deep learning.The beamforming method is used to filter out noise in different directions from the sound source.Finally,the noisy speech in the same direction as the sound source is enhanced by the post-filtering method based on the improved cosh measure to achieve the purpose of speech enhancement.The main work of this thesis is embodied as follows:(1)Research on binaural sound source localization algorithm based on deep learning: First,the binaural speech signal is decomposed into several channels by the Gammatone filter,the high-energy channel is extracted by the weighting coefficient.Then four characteristic parameters representing the spatial information of sound sources,namely,interaural time difference,interaural correlation,interaural level difference and interaural phase difference,are extracted by HRTF.ITD and IC are used as input of the first-level deep learning to obtain the quadrant where the sound source is located.Then the four features are used as the input of the second-level deep learning to obtain the azimuth within the range of 90 degrees.Finally,according to the two the output of the depth learning module determines the precise azimuth of the sound source,achieving a sound source localization of 72 azimuths in a horizontal step of 360 degrees in 5 degree steps.(2)Research on speech enhancement algorithm based on cosh measure filtering:The traditional Wiener filtering method can greatly reduce the energy of speech while eliminating noise.In order to improve the Wiener filter's reduction of speech energy,this paper uses the cosh measure method to improve,using logarithmic form of measure for speech enhancement.Since the frequency and bandwidth of the speech signal exhibit a logarithmic nonlinear relationship,the logarithmic property is more suitable for describing speech signals.(3)Research on directional speech enhancement algorithm in digital hearing aids:The algorithm first uses the speech enhancement algorithm based on sound source localization to obtain the spatial information of the sound source direction,and then uses the beamforming method to remove the noise in different directions from the target speech.The post-filtering based on the improved cosh measure is then used to remove noise in the same direction as the sound source.The experimental results show that the proposed algorithm can enhance the directional speech enhancement of the target speech in a specific direction,so that the enhanced speech improves speech intelligibility and comfort.Firstly,the proposed binaural source localization algorithm can obtain the direction information of the source,which ensures that the beamforming method can remove the speech in different directions from the source.In the direction of sound source,the improved post-filtering speech enhancement algorithm based on cosh measure can further retains speech energy.Compared with the contrast algorithm,the directional speech enhancement algorithm on digital hearing aid has better denoising ability.