Robust Classification Techniques for Acoustic Signal Analysis

Classiication of short duration acoustic signals can be very diicult due to the high degree of variability in the signatures. Input feature vectors, resulting from wavelets or short time Fourier analysis, are typically of high dimensionality, noisy, and contain incomplete information. In this paper, robust artiicial neu-ral networks (ANNs) are identiied that are less sensitive to noisy feature vectors, and that provide a sound estimate of the posterior class probabilities. These classiiers include the \optimumbrain damage"version of the multilayered perceptron and an elliptical basis function classiier. Since diierent classiication techniques have diierent inductive biases, more accurate and robust classiication can obtained by combining the outputs of multiple classiiers. Two approaches to output combination are presented that yield better results for real oceanic signals, and also provide a basis for detecting outliers and \false alarms".