Sixth European Conference on Speech Communication and Technology

This paper describes a Korean utterance verification system based on subword modeling for a vocabulary independent speech recognition system. We deploy strategy consisting of two modules: recognition and verification, for utterance verification. In the stage of recognition, multiple hypotheses with hypothesized word boundaries obtained through Viterbi segmentation of the utterance are obtained. And likelihood ratio is used as a postprocessor for rejecting unlikely hypothesis in the stage of verification. Our study is focused on the verification module. First, we make a comparative study on averaging methods for obtaining the confidence measure for words from the log likelihood ratio based on phone. Three kinds of average techniques were investigated as arithmetic, geometric, and harmonic averages. Second, we study the effect of cohort set, which is the most competitive units to subword units. One cohort set model is trained for each subword. We found out the size of the cohort set for best recognition result. Finally, we present how to model antimodels for each contextdependent units. Three kinds of approaches are studied. The first one is to use cohort set based on contextindependent unit to simplify the calculation. The second one is use cohort set based on contextdependent unit, which is obtained by phone recognizer based on contextdependent units. The final one is to use cohort set based on hybrid units. We make a comparative study on each approach.
Full Paper (PDF) GnuZipped Postscript
Bibliographic reference. Koo, MyoungWan / Lee, SunJeong (1999): "An utterance verification system based on subword modeling for a vocabulary independent speech recognition system", In EUROSPEECH'99, 287290.