Empirical Mode Decomposition for Analysis and Filtering of Speech Signals

Speech signals typically have a stationary interval of 20–30 ms. Due to this, most speech processing techniques split speech signals into segments shorter than the stationary interval to take advantage of the piecewise stationary property of speech. However, there is no way to guarantee that the segments coincide with the stationary timescales inherent in the signal. Furthermore, how do we analyze speech signals over lengths longer than the stationary time scales? Second, there is evidence of the presence of nonlinearities in speech data from the published literature. In this article, the analysis of speech signals, without restriction to stationary time scales, using empirical mode decomposition (EMD) is presented in which the signal is broken down into components called intrinsic mode functions. EMD is especially suited for nonstationary and nonlinear data. The utility of this method, its effects, and opportunities for further research in the context of speech signals are presented.