A stochastic/feature based recogniser and its training algorithm

The authors present a phoneme-based speech recogniser and the training algorithm used to determine the parameters of their model of the speech process. The recognizer uses a speech model which attempts to incorporate the best aspects of stochastic (hidden Markov model) and feature-based approaches. There are two important aspects of the recognizer which distinguish it from others. The first is that the type of parameters used to represent each member of the phone set is phone-class-specific, and the second is the use of a dynamic model of speech parameter movement. The latter enables the authors to represent coarticulation effects more accurately. Thus speech subunits (phones) are divided into six different classes, and front end parameters deemed most appropriate for describing each of these classes are used. Modeling of coarticulation effects is done by working in terms of parameters, including formants, and describing the 'journey' from phone to phone in terms of trajectories to and from targets associated with each phone. >