Classification of transient sonar sounds using perceptually motivated features

This paper describes a novel framework for classifying underwater transient signals recorded by passive sonar. The proposed approach involves two key ideas. Firstly, a feature-selection algorithm is used to identify those acoustic features that optimally model each class of transient sound. Secondly, features that are perceptually motivated are proposed, i.e., they encode information that human listeners are likely to use in transient classification tasks. Three perceptual features are proposed, which encode timbre, the physical material of the sound source, and the temporal context (pattern) in which the transient occurred. The authors show how these features, which are computed over different temporal windows, can be combined to make classification decisions. The performance of the proposed classifier is evaluated on a corpus of transient signals extracted from passive sonar recordings. Specifically, the performance of the perceptual features is compared with spectral features and with those that encode statistics of time, frequency, and power. The present results show that the perceptual features provide valuable cues to the class of a transient. However, the best performing classifier was obtained by selecting a subset of perceptual, spectral, and statistical features in a class-dependent manner.