Current large vocabulary continuous speech recognisers (LVCSR) do not model the effects of speech rate on the speech unit durational characteristics. This paper presents work on the investigation of speech rate, presents three durational models which make use of this rate information and the integration of the models into a TIMIT based LVCSR is described. Although TIMIT contains controlled, read speech, with little speech rate variation, experimental work has shown that the relative duration models produce greater word error rate improvements than models which do not take account of the speech rate.
Bibliographic reference. Jones, M. / Woodland, Phil C. (1993): "Using relative duration in large vocabulary speech recognition", In EUROSPEECH'93, 311-314.