5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

Capturing Discriminative Information Using Multiple Modeling Techniques

Ji Ming, Philip Hanna, Darryl Stewart, Saeed Vaseghi, F. Jack Smith

Queen's University of Belfast, UK

An acoustic model is a simplified mathematical representation of acoustic-phonetic information. The simplifying assumptions inherent to each model entail that it may only be capable of capturing a certain aspect of the available information. An effective combination of different types of model should therefore permit a combined model that can utilize all the information captured by the individual models. This paper reports some preliminary research in combining certain types of acoustic model for speech recognition. In particular, we designed and implemented a single HMM framework, which combines a segment-based modeling technique with the standard HMM technique. The recognition experiments, based on a speaker-independent E-set database, have shown that the combined model has the potential of producing a significantly higher performance than the individual models considered in isolation.

Full Paper

Bibliographic reference.  Ming, Ji / Hanna, Philip / Stewart, Darryl / Vaseghi, Saeed / Smith, F. Jack (1998): "Capturing discriminative information using multiple modeling techniques", In ICSLP-1998, paper 0263.