On the segmentation of narrowly-spaced noisy audio signals

This paper proposes an efficient method of segmenting noisy audio signals using a linear binary Walsh transform, when the signal components are closely spaced and the time intervals between adjacent signal components are unknown. It is shown that the Walsh transform is appropriate for segmenting a noisy waveform. A subset of the Walsh functions is chosen to cover principally the noise subspace such that the resulting linear combination of the selected basis functions captures the features that can discriminate between signal and noise. In the absence of a priori information about the signal and noise statistics, the proposed scheme is based on the linear combination of those basis functions which must be able to identify the adjacent signal components. It is not necessary that the basis functions reconstruct the noise-free versions of the signal components. The only restriction is that the segment length should be some integer power of 2 for the most accurate segmentation. The simulation examples show effectiveness in the segmentation of narrowly separated, noisy signals by using our simple segmentation method.