Spectrum estimation with Gaussian mixture particle filter

Abstract

Based on the idea of using the finite Gaussian mixture model to approach a complex probability distribution, a spectrum estimation algorithm using Gaussian mixture particle filter is proposed in this paper for non-stationary signals. In order to balance the accuracy and the dynamic performance, a revised time-varying autoregressive model is presented combining with Gaussian mixture particle filter, which can estimate the frequency directly and quite reduce the computation cost. Experimental result from the recordings of underwater voice shows that, the proposed method has great performance for non-stationary signals with frequency steps.