Research On Individualization And Spatial Interpolation Methods Of Head-related Transfer Functions

As the key data to researches on spatial hearing,Head-Related Transfer Function describes the filtering effect of the pinna,head and torso of a listener as a sound propagates from the source to the ear drum in free space.Over the years,the reduction of 3-D audio has drawn increasing attention in the field of acoustics with the development of virtual reality technology.Since HRTF contains the directional information of sound sources,it is vital in applications such as modeling indoor sound field,simulating the spatial hearing,etc.HRTF is a spatially continuous function that varies across subjects,because it is filtered by anthropometric features,and the direction of sound source is also continuous in space.The traditional way to obtain HRTF is experimental measurement.However,measurement is time and labor consuming,because it can be conducted only at discrete directions.Specific measurement environment and equipment are also indispensable.Based on that,obtaining HRTFs of specific subject at specific directions fast and efficiently is still to be solved.On the basis of studying the properties of HRTFs,this research focuses on the individual variation and spatial continuity of HRTFs.The main contents are:1)A HRTF database of Chinese subjects is measured and constructed,the properties of HRTFs in the database are also analyzed.The commonly used HRTF databases,such as CIPIC,MIT,etc,are based on the westerners.In the current research,HRTFs at 732 directions of 56 Chinese subjects are measured,50 anthropometric features of each subject are also measured.After processing the original measured data,a HRTF database is conducted.The properties of HRTFs in the database,including time domain property,frequency domain property,interaural time difference,interaural level difference and spectral property,are analyzed.2)Based on the measured HRTF database,statistical methods are used to analyze the relationship between anthropometric features and HRTFs.The reference features that are closely related to HRTFs are selected.Further,three HRTF individualization methods based on reference features are advanced,including database matching method,partial least square regression method and generalized regression neural network method.Spectral distortion is adopted as the criterion to evaluate the precision of the algorithms.The experimental results show that the reference features based methods are effective to obtain the individualized HRTFs.3)HRTF individualization methods based on sparse representation are studied.Considering that the measured HRTF data contains information irrelevant to the direction of sound sources,a HRTF individualization method based on sparse representations of both reference anthropometric features and HRTFs is proposed.Experimental results show that the proposed sparse based method is algorithmically effective in improving the precision of individualization.4)The theory of deep learning is used in the spatial interpolation of HRTFs.An interpolation method of HRTF based on deep neural network is proposed.By adding stacked Restricted Boltzmann Machine to pre-train the net,a deep neural network is constructed for HRTF interpolation.Signal distortion ratio is adopted as the criterion to evaluate the precision of the interpolation algorithm.Experimental results show that the proposed deep neural network based interpolation method is more efficient than the traditional back propagation neural network in HRTF interpolation.5)By using psychoacoustic experimental methods,four subjective listening tests are conducted to evaluate the precision of different HRTF obtaining methods.HRTF assessing methods based on paired comparison and semantic differential test are proposed.Binaural localization test and the least perceptible azimuth test are conducted to assess the performance of different HRTF individualization methods.Binaural localization test and spatial information distinguishable degree test are conducted to assess the performance of different HRTF interpolation methods.The experimental results indicate that individualized HRTFs are available in decrease the least perceptible azimuth effectively when compared with the non-individualized ones,and the deep neural network based interpolated HRTFs are efficient in implementation of HRTF interpolation.