RISC: An Improved Costas Estimator-Predictor Filter Bank For Decomposing Multicomponent Signals

Abstract

We propose an improved version of an estimator-predictor filter bank, originally proposed by Costas [l], for decomposing and tracking multiple, nonstationary sinusoidal components present in a signal. Each component is assigned a signal estimator which is a causal filter, and a predictor. The estimator-predictor combination estimates the next time-sample of its signal component, which is then subtracted from the composite input signal. Ideally, no signal component will then interfere with accurate estimation of the others. However, Costas’s predictor performs poorly when there are components with rapidly changing envelopes. In this paper, we propose an improved predictor that compensates for the group delay introduced in the signal components by the causal filtering, by minimizing a prediction error criterion. With this improved predictor, using a computer synthesized multicomponent signal, we show that we achieve cleaner separation of signal components when compared with Costas’s method. We also show that this method can be used to separate the essentially harmonic partials in voiced speech.