A STUDY ON ROBUST UTTERANCE VERIFICATION FOR CONNECTED DIGITS RECOGNITION

Utterance verification represents a key technology in the design of a user-friendly speech recognition system. One essential element when de signing such a system is the ability to maintain a uniform performance over a wide range of acoustic conditions. An acoustic mismatch betwee n training and testing conditions often results in an undesirable perf ormance degradation This paper addresses the issue of robustness in ut terance verification of a speech recognition system. Two techniques, n amely signal bias removal (SBR) and on-line adaptation, are studied. T he SBR algorithm is used to deal with global mismatch conditions cause d by handset and channel differences. The on-line adaptation algorithm is used to adjust verification threshold at runtime for achieving a d esirable trade-off between false rejection and false alarm in new test conditions. Various on-line adaptation schemes are investigated. We s how that both supervised or unsupervised adaptation can effectively ad just the verification threshold to achieve a desirable performance tra de-off irrespective of the initial setting of the threshold. We report on connected digit recognition/verification results for matched and m ismatched training and testing conditions. At a 5% digit string reject ion rate, the proposed robust utterance verification system gives a re duction in string error rate between 32% and 35% over the conventional system, while still correctly rejects over 99.9% of nonvocabulary utt erances. (C) 1997 American Institute of Physics.