(An) out-of-vocabulary rejection technique using anti-phoneme model and on-line garbage model in hidden Markov model

Abstract

Automatic Speech Recognition (ASR) has made great advances in last 10 years. The application of ASR in everyday situation represents a new possibility in human-machine interface. For this to become a reality, it is important to distinguish in-vocabulary words from out-of-vocabulary words. In order to reject out-of-vocabulary words effectively while accepting in-vocabulary words, conventional approaches such as the filler model approach or the on-line garbage model approach have been proposed. These approaches either require the use of extraneous data to train filler models or require adjusting when the set of in-vocabulary words is changed. In this thesis, a novel approach based on probabilistic characteristics is proposed to reduce the confusion between a claimed model and the other models. In order to reflect probabilistic characteristics of models, the anti-model for the claimed model is constructed by weighting observation probabilities of other models with their weights being inversely proportional to their distances to the claimed model. In addition, a hybrid of the proposed model and the on-line garbage model is also suggested to improve performance. The proposed method is evaluated using 455 Korean isolated words speech corpus. For simulation, 90 words are selected as in-vocabulary words and the same number of words are selected as out-of-vocabulary words without overlapping. The proposed method results in 8.33% of equal error rate, which is improved by 63.16% for the filler model approach and 39.77% for the on-line garbage model in error rate reduction.