8th European Conference on Speech Communication and Technology

Geneva, Switzerland
September 1-4, 2003


An Efficient, Fast Matching Approach Using Posterior Probability Estimates in Speech Recognition

Sherif Abdou, Michael S. Scordilis

University of Miami, USA

Acoustic fast matching is an effective technique to accelerate the search process in large vocabulary continuous speech recognition. This paper introduces a novel fast matching method. This method is based on the evaluation of future posterior probabilities for a look-ahead number of timeframes in order to exclude unlikely phone models as early as possible during the search. In contrast to the likelihood scores used by more traditional fast matching methods these posterior probabilities are more discriminative by nature as they sum up to unity over all the possible models. By applying the proposed method we managed to reduce by 66% the decoding time consumed in our time-synchronous Viterbi decoder for a recognition task based on the Wall Street Journal database with virtually no additional decoding errors.

Full Paper

Bibliographic reference.  Abdou, Sherif / Scordilis, Michael S. (2003): "An efficient, fast matching approach using posterior probability estimates in speech recognition", In EUROSPEECH-2003, 1161-1164.