Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Speech Recognition Using a Linear Dynamic Segmental HMM

Wendy J. Holmes, Martin J. Russell

Speech Research Unit, DRA Malvern, Malvern, Worcs., UK

This paper describes research into linear-trajectory dynamic segmental hidden Markov models (HMMs). The main advantage of these models over conventional HMMs is that they allow explicit modelling of speech segment dynamics. In general terms, a trajectory-based segmental HMM provides a parametric representation of the range of possible underlying trajectories for a speech sound. Acoustic feature vectors are regarded as noisy observations of a particular trajectory. This model represents an extension and generalization of the previously-developed static segmental HMM [1], which can be viewed as a constant-trajectory model. In the present paper, a linear-trajectory segmental HMM is described and Baum-Welch-type re-estimation formulae are presented. Preliminary recognition experiments on a connected-digit recognition task have demonstrated performance improvements over the previous results with static segmental HMMs, which in turn outperformed conventional HMMs.

Full Paper

Bibliographic reference.  Holmes, Wendy J. / Russell, Martin J. (1995): "Speech recognition using a linear dynamic segmental HMM", In EUROSPEECH-1995, 1611-1614.