Using artificial neural networks to model the human annoyance to sound

Peter Collom Laux, Purdue University

Abstract

There are currently no models for the human perception of the annoyance of noise that give accurate predictions for all types of noise stimuli. An Artificial Neural Network (ANN) to model Human annoyance to noise was developed. There is a practical necessity to minimize the size of the network structure because of the fact that the number of data points required to train an artificial neural network is related to the number of variables in the network. Since the output training data for this network are the average human responses to sounds, only a small number of training data sets were available. A minimally sized ANN architecture was developed to deal with the lack of training data, and this network was initialized by training it to emulate Zwicker's Unbiased Annoyance model. This approach to initializing the network was more successful that assigning random numbers to variables. The model was then trained to the average of 28 subjects responses to noise stimuli with varying: modulation envelope shape, modulation frequency, maximum level, modulating function, frequency band for the noise, type of added tones, and level of added tones. An extension to the model to include roughness as an input, in addition to Percentile Loudness, Sharpness, and Fluctuation Strength was developed. Re-training this expanded artificial neural network model produced a model for annoyance that accurately predicts the annoyance ratings for a validation set of signals, not used for the network training. Future model expansions could include incorporating other measures of signal quality, e.g., tonality as an input. Approaches to expanding the network to do this are discussed.

Degree

Ph.D.

Advisors

Davies, Purdue University.

Subject Area

Mechanical engineering|Acoustics|Audiology

COinS