In this paper, a fast confidence measure algorithm for continuous speech recognition is presented. The posterior probabilities of states are selected as the main feature. The confidences of hypothesized phonemes and words are estimated using the information of corresponding states. Traditional systems use two passes and two different models for decoding and calculating confidence respectively. In this new algorithm, the confidence score is calculated during the decoding of the state graph constructed from task grammar. And in this new algorithm, the same acoustics model is used during the process of decoding and confidence estimation. The old and new computing algorithms are compared through experiments, and the result shows that performance of the new algorithm is effectively improved. The equal error rate (EER) of new fast confidence algorithm is reduced and the speed is accelerated.
Cite as: Dong, B., Zhao, Q., Yan, Y. (2005) Fast confidence measure algorithm for continuous speech recognition. Proc. Interspeech 2005, 1457-1460, doi: 10.21437/Interspeech.2005-516
@inproceedings{dong05_interspeech,
author={Bin Dong and Qingwei Zhao and Yonghong Yan},
title={{Fast confidence measure algorithm for continuous speech recognition}},
year=2005,
booktitle={Proc. Interspeech 2005},
pages={1457--1460},
doi={10.21437/Interspeech.2005-516}
}