This paper deals with the optimization of a confidence measure based on HMM rescoring of the hypothesized word. This confidence measure is expected to be very cheap. It is only based on the recognized hypothesis and the HMM responses. Neither anti-models nor N-Best hypothesis are used. This study investigates 2 issues: one is the dispersion of the HMM score among the phonemes, the other is a discriminant weighting of the acoustical features. This confidence measure is evaluated in the framework of a large vocabulary directory attendant application. Evaluation shows that exploiting the dispersion of the response scores is a promising approach, whereas the weighting does not give major improvement. These findings confirm that the acoustical parameterization used is suitable for this task.
Cite as: Charlet, D. (2000) Optimizing confidence measure based on HMM acoustical rescoring. Proc. ASR2000 - Automatic Speech Recognition: Challenges for the New Millenium, 203-206
@inproceedings{charlet00_asr,
author={Delphine Charlet},
title={{Optimizing confidence measure based on HMM acoustical rescoring}},
year=2000,
booktitle={Proc. ASR2000 - Automatic Speech Recognition: Challenges for the New Millenium},
pages={203--206}
}