Linear prediction based on Teager-Kaiser energy function and application to speech modeling and spectral analysis

In this paper, the classic problem of determining the optimal linear FIR filter that minimizes an appropriate norm of the filter output is addressed, with the signal norm being the recently introduced Teager-Kaiser Energy norm. Normal equations are derived and conditions for the existence of a solution are given. This theory is then applied to LPC modeling of speech signals and compared with the standard LPC results, where the L/sup 2/-norm of the filter output is minimized. It is observed that the TK-LPC spectra have enhanced formant structure and expanded bandwidth, which could be useful in increasing the perceptual quality of TK-LPC coders. Finally, TK-LPC analysis was applied to spectral estimation, where it is observed that the TK-LPC spectra had higher resolution.