Natural number recognition using MCE trained inter-word context dependent acoustic models

Among applications that require number recognition, the focus has largely been on connected digit recognizers. In this paper, we introduce an acoustic model topology for natural number recognition by using minimum classification error (MCE) training of inter-word context dependent models of the head-body-tail (HBT) type. Experimental results on natural number applications involving dollar amounts and U.S. telephone numbers show that using HBT models for natural number data reduces string error rates by as much as 25% over context independent whole word models. In addition, for speech input which is strictly of connected digit type, the increase in string error rates is negligible when a natural number telephone grammar is used instead of a connected digit telephone grammar. This will enable natural number speech recognition systems to be more widely accepted because recognition accuracy is maintained while permitting a more natural and flexible user interface.