A New Lpc Error Criterion for Improved Pitch Tracking

A NEW LPC ERROR CRITERION FOR IMPROVED PITCH TRACKING Mohammad R. Zad-Issa and Peter Kabal Electrical Engineering, McGill University, Montreal, Quebec, Canada H3A 2A7 ABSTRACT In Linear Predictive coders the output of the LP analysis lter is used to represent the glottal excitation signal. For high pitched voices during nasal sounds or nasalized vowels, the speech signal takes on a sinusoidal shape. The corresponding residual signal has a very low energy and the pitch pulses are weak or absent, resulting in poor pitch tracking. These segments of speech are also characterized by large frame-to-frame variations of the LP coe cients. In this paper we propose a composite formant prediction error criterion leading to a clear track of residual pulses even for for the sinusoid-like speech, while enhancing the smoothness of the lter parameter evolution.