Improving subband spectral estimation using modified AR model

It has already been shown that spectral estimation can be improved when applied to subband outputs of an adapted filterbank rather than to the original fullband signal. In the present paper, this procedure is applied jointly to a novel predictive autoregressive (AR) model. The model exploits time-shifting and is therefore referred to as time-shift AR (TS-AR) model. Estimators are proposed for the unknown TS-AR parameters and the spectrum of the observed signal. The TS-AR model yields improved spectrum estimation by taking advantage of the correlation between subseries that arises after decimation. Simulation results on signals with continuous and line spectra that demonstrate the performance of the proposed method are provided.