A alternative way of representing time variations of the speech spectra is presented. We propose to model the trajectories of the poles of the LPC analysis spectra using exponential functions as alternative to delta parametes. The obtained time constants are incorporated in the observation vector and used for recogniton in an HMM based recognition system. The performance is evaluated on a database consisting of connected digits. The recognition results are compared with results obtained using delta and acceleration coefficients.

Peer Reviewed