Sixth International Conference on Spoken Language Processing
(ICSLP 2000)

Beijing, China
October 16-20, 2000

Utterance Verification Based Speech Recognition System

Beng Tiong Tan, Yong Gu, Trevor Thomas

Vocalis Ltd, Cambridge, UK

Many existing search algorithms aim at searching for the best hypothesis from all possible hypotheses with the help of techniques like beam search to reduce the computational cost. These search algorithms are based on the competitive criteria because the best hypothesis is determined after we have the knowledge of all other possible hypotheses. In this paper, we investigate the possible use of a qualifier criterion. A qualifierbased search will make recognition decision without considering all possible hypotheses. Instead of finding the best hypothesis, the system will try to find a hypothesis that can meet with some criteria reflecting the goodness of a hypothesis. We investigate the use of utterance verification measurement as the qualifier measurement used in the search. An utterance verification based speech recognition (UVSR) system is proposed. The search will be terminated as soon as the qualifier criteria are meet without the need to evaluate the likelihoods of all possible hypotheses. The proposed search algorithm with UV performs very closely to the standard speech recognition with UV but with 23% less in the average searching time. If the information of language model is provided, the average searching time can be reduced by 43%.

Full Paper

Bibliographic reference.  Tan, Beng Tiong / Gu, Yong / Thomas, Trevor (2000): "Utterance verification based speech recognition system", In ICSLP-2000, vol.2, 899-902.