Acoustic Quality Parameters Used for Error Evaluation of Neural Networks Modeling for HRIRs Applied to Escape Training in Blind Conditions

This work presents an accuracy analysis of a system for Head-Related Impulse Responses interpolation based on artificial neural networks (ANN). The error analysis behavior in the time domain, between the target and the output functions, could lead to mistaken results if the functions were slightly time-shifted. On the other hand, frequency domain errors may be influenced, for instance, by high frequency components, where the resolution is smaller than at lower frequencies and at frequency ranges that does not interfere in the human being perception. The proposed criteria are based on adequating well known room acoustic quality parameters for evaluating the interpolation error. The article presents the ANN architecture and the optimal configuration (in terms of processing speed and accuracy) based on the proposed error criteria. The comparative results were obtained from the receiving area where human auditive capability is most sensible and the interaural differences are most notorious (horizontal plane). The results show that the modified acoustic parameters provide a good estimation of HRIR interpolation comparisons.