For the recognition of words by a speech recognition system a description of particular words as a part of the reference knowledge is ncessary, which describes the real appearence of the words as real as possible. Depending on the concrete recognition task, that is the recognition of single words, of controlled fluent spoken word strings (read speech) or of unreglemented spontaneous speech, there are different problems of modelling. In the case of single word recognition we have mostly complete and clear articulation, while by the realisation of word strings or spontaneous speech we find an increasing variation of articulation, implied by speech economy. Here we find changes at the word level and for spontaneous speech there are changes over word boundaries too. The present paper deals with the investigation of read and spontaneous spoken speech and with the description of the observed phenomena. We derive rules for the description of pronunciation variants and we propose structures for pronunciation dictionaries as a part of the reference knowledge, which represent a powerfull modelling. The proposals are verified by recognition experiments.
Bibliographic reference. Flach, Gudrun (1995): "Modelling pronunciation variability for special domains", In EUROSPEECH-1995, 1743-1746.