In this paper a new approach for modelling time variations in the speech spectra is presented. We propose to approximate the trajectories of the frequency and amplitude of the poles of the LPC spectra with exponential functions. The obtained time constants of the exponential functions are incorporated in the observation vector and used for recognition in an HMM based recognition system. The performance of the new parameters is tested using a database of connected digits. The recognition rates obtained with the new parameters are compared to results obtained with delta parameters.

Peer Reviewed