Optimized Learning of Spatial-Fourier Representations from Fast HRIR Recordings

The acquisition of head-related impulse responses (HRIRs) has traditionally been a time-consuming acoustic measurement process. Novel continuous-azimuth recording techniques have dramatically accelerated the acquisition, but conversion into continuous Spatial-Fourier representations (SpaFoR) of HRIRs provides a host of cumbersome implementation challenges. The direct closed-form least-squares approach is unfortunately not practical and we will therefore explore the retrieval of SpaFoR model parameters of HRIR by contemporary machine-learning tools. Specifically, we employ the standard stochastic-gradient learning with Tensorflow on a graphics processing unit (GPU) and compare its performance with previous covariance-based least-squares on the general purpose processor. Apart from the sought simplification and acceleration, our paper is dedicated to hyperparameter optimization in order to make sure the final state of the machine learning approach still attains the accuracy of the optimal least-squares solution. The paper finally applies the proposed method to a real acoustic HRIR recording to illustrate the validity of the system identification obtained by learning.