Some Preliminary Results on Speaker Independent Recognition of the DARPA Resource Management Task (NOT PUBLISHED)

In this paper, we report on some preliminary speaker independent speech recognition results obtained on the DARPA resource management task. A series of recognition experiments was conducted to determine the extent to which standard modeling techniques for continuous density hidden Markov models (CDHMM) could be applied to this problem. Our initial studies focused mainly on the selection of a good set of fundamental speech units to characterize both the acoustic and phonetic nature of continuous speech.