A new hybrid decoding algorithm for speech recognition and utterance verification

We propose a new search algorithm for one-pass utterance verification in which verification and recognition are simultaneously done. The basic idea is that we make a hybrid decoder in which the conventional Viterbi decoder is combined with a likelihood ratio decoder based on a confidence score representing the confidence level for utterance verification. We overcome the dynamic range problem encountered in the likelihood ratio decoder by use of a sigmoid limiter. We also improve its robustness by selective use of the anti-phone models in calculating the phone likelihood ratios. The proposed strategy is applied to a 1000-word Car Reservation task. Experimental results show that the proposed hybrid decoder gives better results than the convention likelihood decoder or the likelihood ratio decoder, particularly in dealing with out-of-vocabulary words or out-of-task sentences.