Interpolation of Minimum-Phase HRIRs Using RBF Artificial Neural Network

Now head-related transfer function (HRTF) databases we already have are not spatially continuous, it's necessary to reconstruct a high spatial resolution HRTF database to solve this problem. Based on the minimum-phase HRIRs (Head-related Impulse Responses), we analyze and compare the errors of three traditional methods include linear, cubic, and spline interpolation. Furthermore, we propose a method using RBF (radial basis function) artificial neural network to interpolate the minimum-phase HRIRs. The minimum-phase reconstructed empirical HRIRs are used to train and establish the neural network. By training RBF neural network, we can approximate HRIRs in any spatial positions. The experimental results show that this method retains the advantages of minimum phase HRIR. It has the minimum group delay, minimum filter length, less interpolation error and a good performance for estimation. The proposed method makes more convenient to obtain HRTFs at any required spatial positions in synthesizing virtual 3D sound.