Redes neurais artificiais aplicadas à modelagem de respostas impulsivas associadas à cabeça humana para gerar aurilização

Abstract

One of the objectives of the development and improvement of techniques in systems of generation of aurilization consists in the reduction of computational cost and to guarantee the quality of acoustic simulation. In order to obtain the auralization, it is necessary, besides an anechoic sound, the determination of the binaural room impulsive responses (BRIRs), for a determined pair of source-receiver. In this sense, one way to reduce the complexity of such systems is to reduce the computational cost of calculating the spectral modification of the head-related impulse responses (HRIRs). For this, in this work, the RBF neural networks were used to replace the traditional method. To cover the entire auditory reception space, a set of artificial neural networks was adopted, each network acting in a specific direction. Each neural network has as input the acoustic ray s spectrum in octave bands and as output the filtered HRIR of the corresponding direction. The strategy of training these neural networks, the implementation of the networks, the model and the observations are presented. By means of the proposed method, a computational cost reduction of approximately 70% was obtained in relation to the traditional method, which represents a significant reduction in the implementation of the auralization systems. To verify the efficiency of the approach, several comparisons were made between the filtered HRIRs using the RBF neural networks and the traditional method. The similarity between the results was evaluated from the normalized cross-correlation coefficient. The simulated results demonstrate the efficiency of the technique using the neural networks, since these have values of correlations close to one. Only a discrepancy in some phase components of HRTFs has been observed for some directions, however, these errors are unnoticeable. Also presented are some results of the international intercomparison of room acoustics simulators, called Round Robin on Auralization. For the calculation of the BRIRs, the code RAIOS (Room Acoustics Integrated and Optimized Software) and the RBF neural networks were used. The simulated and measured results are compared and discussed.