Context dependent anti subword modeling for utterance verification

Utterance verification is used in spoken language dialog systems to reject the speech that does not belong to the task and to correctly recognize the sentences that do. Current verification systems use context dependent (CD) or context independent (CI) subword models and CI anti-subword models. We propose many methods of modeling the CD anti-subword models. We have compared these anti-models and show that the anti-models with the same context have the most separation between the speech that contains the subword and the speech that does not contain the subword. We have also conducted recognition/verification experiments with a two pass verifier and two one pass verification systems to compare the different types of anti-subword models. Our results show that the same context anti-subword models have the best recognition/verification performance.