Linear modeling algorithm for tracking time-varying signals

This paper presents a new algorithm for tracking the spectrum of non- stationary signals. In general there is no law relating frequency and time, and therefore, the frequency-time curves are usually approach dependent. The algorithm described here is an extension of the well-known Levinson model for estimating the spectra of stationary signals. The signal parameters are estimated by fitting the model with time-varying coefficients based on an exponential forgetting factor that is introduced to the autocorrelation function. The first operation is the excitation with the input sequence y(n), n = 0, 1, 2, ..., N, to produce a scalar output, then time-updating by incrementing the previous value with a scalar. To demonstrate the effectiveness of the algorithm, some numerical examples are considered: chirp signal in white noise, two sinusoids, and speech signals.