Room acoustics modeling using a hybrid method with fast auralization with artificial neural network techniques
* Presenting author
One of the objectives of the development of numerical techniques in acoustic virtual reality systems and the production of reliable and plausible auralizations is to reduce the computational cost and, simultaneously, to guarantee the sound quality. In this article, a new technique is presented to model the head-related impulse responses and its filtered versions, necessary for the binaural room impulse responses computation. Artificial neural networks of the radial basis functions kind are used. A set of such networks is trained and tested to cover the entire auditory space around the human head. Each neural network is associated to a given direction and has, as input, the power spectrum of the sound wavefront in octave bands that reaches the receiver and, as output, the filtered head-related impulse response, in the corresponding direction. With this strategy, the computation is performed directly in time domain, bypassing the need of convolving each wavefront with the head-related impulse response for the considered direction. As it is shown, a reduction of the computational cost of about 85% is achieved. The result obtained with the proposed method is compared with that computed with the classical convolution method, both in time and frequency domain showing a negligible difference.