This paper describes a new approach to im- prove the discriminative capabilities of poorly trained Hidden Markov Models. A discriminative post-classifier is developed for the recognition procedure, which reduces the classification to pairwise phoneme comparisons by weighting the components of the feature vectors according to their ability to separate the phoneme pairs. In this paper the application to single word and continuous speech recognition is presented. The discriminative approach has much effect on models, trained with a small database. In experiments the speaker independent recognition rate of German digits could be increased from 90 % to 96 %. With a larger training database and well trained models, the recognition rate of 99. 5 % can not be further improved. Keywords: Hidden Markov Model, discriminative recognition, transinformation, Kolmogorov distance
Bibliographic reference. Zünkler, Klaus (1991): "A discriminative recognizer for isolated and continuous speech using statistical separability measures", In EUROSPEECH-1991, 1095-1098.