For speech recognition systems that use stochastic models it is usual to choose the model topology independently of the phonetic structure of the recognition units. Such an arrangement can cause difficulties for modelling linguistic units that clearly have alternative forms of phonetic realization. Choosing a general topology which allows for alternatives gives undesirable tolerance for those units where such variations are not expected. This paper presents a design and training technique which selects the topology for each unit to suit its expected phonetic structure. A recognizer using these techniques is described. Keywords: Speech recognition, stochastic models, phonetic knowledge, forward-backward algorithm
Bibliographic reference. Holmes, John N. (1991): "Use of phonetic knowledge when designing and training stochastic models for speech recognition", In EUROSPEECH-1991, 1257-1260.