This paper discusses issues affecting the design of regression features for use in word recognition. Particular attention is paid to interactions with environmental noise effects, including additive noise and the Lombard reflex. We discuss separate control of the time length and number of frames used to calculate the regression features, selection of regression window parameter values, and the selection of regression feature order (up to third order). Recognition results are shown on a vocabulary of confusable English alpha-digits and function words, and on the JEIDA Japanese Cityname vocabulary. KEYWORDS: speech recognition, regression feature, delta cepstrum, Lombard, noise
Bibliographic reference. Applebaum, Ted H. / Hanson, Brian A. (1991): "Tradeoffs in the design of regression features for word recognition", In EUROSPEECH-1991, 1203-1206.