Nonnegative features of spectro-temporal sounds for classification

A parts-based representation is a way of understanding object recognition in the brain. The nonnegative matrix factorization (NMF) is an algorithm which is able to learn a parts-based representation by allowing only non-subtractive combinations (Lee and Seung, 1999). In this paper we incorporate a parts-based representation of spectro-temporal sounds into the acoustic feature extraction, which leads to nonnegative features. We present a method of inferring encoding variables in the framework of NMF and show that the method produces robust acoustic features in the presence of noise in the task of general sound classification. Experimental results confirm that the proposed feature extraction method improves the classification performance, especially in the presence of noise, compared to independent component analysis (ICA) which produces holistic features.