2D semi-NMF of scale-frequency map for environmental sound classification

Abstract

This paper introduces a novel two dimensional feature extraction method for environmental sound classification, based on two dimensional semi-nonnegative matrix factorization (2D Semi-NMF) of scale-frequency maps. We first extract scale-frequency maps (SFMs) from the input signals, and this feature is considered preserving scale and frequency characteristics of signals. Second, a 2D Semi-NMF method is applied on SFMs to get more information of the input signals. We use the combinational coefficients extracted from 2D Semi-NMF for classification. Experimental results on an 8 class environmental sound database show that 2D Semi-NMF has better classification accuracy than traditional ID NMF and 2D NMF Also, applying 2D Semi-NMF on SFMs will get slightly improvement than SFMs features alone.