On the use of artificial neural networks to model sound quality

Abstract

Loudness, percentile loudness, sharpness, and fluctuation strength have been shown in laboratory experiments to be correlated with annoyance. Zwicker proposed a model of unbiased annoyance (UBA) based on time of day (d), fluctuation strength (F), sharpness (S), and N10. In psychoacoustic tests on amplitude-modulated noise signals [P. Laux and P. Davies, NCEJ 40 (3) (May-June 1993)] UBA has been shown to have a higher correlation to annoyance than other commonly used noise measures, although, other correlation models were found to produce an even better model of the relationship between (S,N10,F) and the subjective ratings. The use of artificial neural networks (ANNs) to describe the relationship between (S,N10,F) and the subjective responses has been investigated. The smaller the network, the fewer the number of weights, and the fewer the number of signals that a subject has to rate. The objective of the research was to estimate a network that was the smallest possible model of the data. As a starting point UBA was modeled as a function of (S,N10,F); the resulting ANN model of UBA was then used as the initial conditions when subjective data were available.