Modelling the Excitation Function to Improve Quality in LPC's Resynthesis

Celso Aguiar

CCRMA - Center for Computer Research in Music and Acoustics
Stanford University
Stanford, CA 94305-8180 USA


LPC (Linear Predictive Coding) is a well known technic for speech analysis- synthesis. The analysis consists in finding a time-based series of n-pole IIR filters whose coefficients better adapt to the formants of a speech signal. These computations produce a residual signal which is the exact complement to the infomation kept in the coefficients, if we wish to recover the original. The model of the human vocal tract mechanism assumed by LPC presupposes that speech can be reduced to a succession of voiced or unvoiced sounds. Thus, an excitation function composed of pulse or noise substitutes the residual in the resynthesis. This assumption is adequate for some tasks but, in a musical context, the artifacts introduced can yield unsatisfactory results. This article proposes a simple alteration in the model to improve the quality of LPC's resynthesis. Some of the conventional problems like "buzzy" quality and loss of coloration are partially corrected.