The WRLS algorithm for speech processing

Presents a theoretical analysis of the source of error of the weighted recursive least squares (WRLS) algorithm when it is used as a linear predictor. It has been generally accepted that the estimate from the WRLS algorithm has only two sources of error. When the WRLS algorithm is applied to a voiced speech signal, however, another source of error is found to exist. This error is termed the excitation error, since it occurs whenever the input signal has impulsive components. The properties of this excitation error are addressed. The effect of the impulsive components on the tracking ability of the WRLS algorithm is also analyzed, and it is shown that the existence of impulsive components actually enhances the tracking ability of the WRLS algorithm by reducing the initial magnitude of the lag error at epoch locations. This theoretical analysis has then been extended to the implementation of pitch determination algorithm and its verification.<>