Confidence measure based on forced-alignment for out-of-vocabulary term detection

In this paper, we propose a method of confidence measure (CM) to improve the performance of out-ofvocabulary (OOV) term detection. For hypothesized OOV terms, the proposed method firstly obtains the acoustic likelihood by forced-alignment, and then computes the final CM for verification. The forcedalignment provides the mapping relation between the frame observations and the states of hidden Markov models. The CM can be calculated by averaging phone-level confidence or classifying confidence features of syllable with support vector machine. The confidence features take advantage of the merit of Chinese syllable structure and describe the confidences in every sub-syllable level. The experiments conducted on the Hub-4NE Mandarin database show that the proposed method of confidence measure can achieve improvements over the current lattice-based method for OOV detection.