Speech dynamic feature are routinely used in current speech recognition systems in combination with short-term (static) spectral features. The aim of this paper is to propose a method to automatically estimate the optimum ponderation of static and dynamic features in a speech recognition system. The recognition system considered in this paper is based on Continuous-Density Hidden Markov Modelling (CDHMM), widely used in speech recognition. Our approach consists basically in 1) adding two new parameters for each state of each model that weight both kinds of speech features, and 2) estimating those parameters by means of a discriminative training algorithm that minimizes the recognition error using the recently proposed Generalized Probabilistic Descent (GPD) method. Experimental results in speaker independent digit recognition show an important increase of recognition accuracy.

Peer Reviewed