Unknown utterance rejection using likelihood normalization based on syllable recognition

This paper proposed and evaluates a method to reject the unknown utterance based on the likelihood with a normalization based on the likelihood utilizing the syllabe likelihood. In the proposed method, the task and syllabe recognition are executed in parallel, and the normalization for the likelihood is determined as the difference of the two obtained kikelihoods. The the rejection decision is made based on the threshold. By this method, the rejection can be decided by the set threshold, independently of the speaker and the noise conditions