Non-stationary signal analysis using temporal clustering

We present a model of non-stationary time series generated by switching between a finite number of random processes and apply temporal clustering to estimate the model's parameters. Applications of the algorithm to segmentation of non-stationary time series and a simple example of pre-processing a speech signal will be discussed. The model defines a non-stationary composite source generated by randomly switching between elements of a finite number of random processes. The switching probability distribution which underlies the behavior of the switch is controlled by a time varying vector of parameters which is used to determine a different switching probability in each time instant. This definition allows us to analyze a drift between disjoint states of the composite model.