Speech analysis by homomorphonic prediction

Abstract

Linear prediction (LPC) is a generally accepted method for obtaining all-pole speech representations. However, in many situations (e.g., nasalization studies) spectral zeros are important and a more general modeling procedure is required. Unfortunately the need for pitch synchronization has limited the success of available techniques. This talk deals with a new approach to pole-zero identification, based on homomorphic prediction, which avoids the synchronization problem. Homomorphonic prediction is a three step process. Homomorphic deconvolution is used to obtain a minimum-phase estimate of the vocal tract impulse response. Such a signal by definition is properly time synchronized. Linear prediction is applied to this waveform to identify its poles. Next, the LPC “residual” (error signal) is computed by inverse filtering. This signal contains the information about the zeros. Its z transform is approximated by a polynomial either through a weighted least-squares procedure (as in Shanks' method) or by spectral inversion followed by a second evaluation on real speech data will be presented. [Supported in part by ARPA, NSF, and the Fannie and John Hertz Fonudation.]