Secondary testing techniques for word recognition in continuous speech

In the usual pattern matching approach to automatic recognition of speech, some measurement of pattern similarity is used to obtain a tentative word identification. This paper presents techniques for using additional information for further testing of this hypothesis. This information is in the form of other types of spectral match statistics, energy ratio statistics, and energy derivative statistics. A stepwise linear discriminant analysis procedure is used to select, for each template, a subset of the available statistics, and derive a linear combination of these statistics. The resulting score is the log odds in favor of this tentative identification given the observed values of these additional statistics.