Abstract: This paper describes a system of small-vocabulary Mand arin keyword spotting speech recognition based on discrete HMMs. A novel confide nce measure for utterance verification is also presented. The confidence score i s obtained by rescoring putative keywords detected by the keyword spotter via a Viterbi search algorithm on the isolated keyword HMMs. Experimental results show that the confidence measure has good performance on discriminating keyword detections from false alarms.

Key words: speech recognition, keyword spotting, HMM, confidence score, rejection