Automatic feature identification using multidimensional scaling data

Abstract

Multidimensional scaling (MDS) listening experiments have long been used to better understand the perceptual cues human use when distinguishing between sounds in a given dataset. By collecting aural similarity measures between sound pairings, one can find the position of each sound in a D-dimensional perceptual space. The space allows the researcher to explore the perceptual attributes used to distinguish sounds. Typically, candidate signal features are correlated to each dimension to find which features accurately predict the signal positions in the perceptual space. Signal features generally exhibit varying degrees of correlation, with the optimal (highest correlated) feature declared to be the one used in perception. This method is not only ambiguous as to which feature (or combination of feature) is important to perception, but also is limited to previously defined signal processing features. Instead of optimizing over a limited set of features, one can optimize over a function which maps our signals to the perceptual MDS results. This approach allows one to find a transform of the data that better represents the complicated perceptual space. A number of solutions to the problem will be presented based on the results of an MDS experiment using active sonar echoes.