Abstract:

In this talk, a procedure is presented for the development of an artificial neural network model of human annoyance to noise. This model development is based on using available sound quality metrics (loudness, sharpness, fluctuation strength, and roughness) that were calculated using post-process temporal filtering and numerical formulation operations on measured time-varying critical band loudness data obtained from a Bruel and Kjaer loudness analyzer. The artificial neural network (ANN) was initially developed to model Zwicker's equation for unbiased annoyance (UBA). Subsequent training of the artificial neural network was done by using a series of 700 noise signals that were evaluated for their annoyance by a pool of normal-hearing young adult U.S. born subjects in a parameter estimation subjective test procedure. The noise stimuli were varied in seven different properties, which included: peak loudness, modulation frequency, modulation envelope shape, modulation depth, frequency band of modulated noise, addition of continuous complex tone structures, and the relative level of complex tone set to the underlying noise. The response data were averaged across subjects and the UBA-ANN model was retrained using the average subjective responses. The results of this new ANN--annoyance model are compared to other metrics for sound measurement/rating.