Utterance verification On DTW based speech recognition using likelihood

Utterance verification provides the speech recognition system a user-friendly interaction. However, small-vocabulary system using DTW algorithm cannot afford HMM based utterance verification. So, to equip the DTW based recognizer with effective utterance verification becomes an essential problem in a low computational application. We proposed a new utterance verification method, combining both statistic and distance measure capability, to map DTW distance to a certain likelihood. The likelihood as a confidence measure performs a good ability of both speech recognition and utterance verification. With a test set of fifteen words, at 5.24% false rejection, the verification method brought on 10.44% false alarm rate and 93.61% accuracy. Furthermore, 94.76% out-of-vocabulary utterances were correctly rejected.