The ability of FIR neural networks to model subjects' judgments of real-world sounds is investigated. The samples for training the networks are 27 environmental sounds from the natural environment, traffic, house appliances, and the industry. 31 Subjects were asked to rate the sounds by 9 attributes on 5-point category scales. Differences between judgments on known and unknown sounds were quantified by t-tests. The largest difference is about 1 category for the attribute 'annoying'. Due to this fact only judgements on known sounds were taken into account in the following investigation. Factor analysis, performed on the subjects' answers, extracted 2 factors. The factors can be identified to represent the attributes 'annoying' and 'powerful'. In order to reduce the amount of input values for the network a factor analysis was applied to one-third octave band spectra of the sounds. FIR neural networks were trained to predict the factor 'annoying' of the subjects' answers from the factor values as emerged from the analysis on the frequency bands. At the optimum the networks on average produce 3.2% error on the training set and 13.48 error on the validation set. The results are very promising for the application of FIR neural networks for the prediction of sound quality judgments. [References: 6]
展开▼
