An efficient nonkeyword rejection scheme for Mandarin speech recognition

During the decade of 1990's, many researchers have tried to make automatic speech recognition (ASR) systems available in the real world and there are some systems operating already. However there are still some problems which need to be solved, especially for spontaneous speech input. One of the typical problems is to reject utterance that does not include any valid keyword or to verify the keyword embedded in the input utterance. Conventionally, the decision to reject or accept an utterance as keyword is to compare an unnormalized score with a threshold. Recently, some schemes based on normalized scores are proposed to cope with this problem. Furthermore, if has been shown that the methods based on normalized score many provide better performance than unnormalized ones. However the key point of this kind of method is to determine the ratio of two likelihoods which are obtained indirectly from the output of a recognition system and its corresponding antimodel, respectively. In this paper, we will propose an efficient approach to deal with this problem.