Publication Details
Measurement of Complementarity of Recognition Systems
speech recognition, system combination, system complementarity, complementarity measure
Combination of different speech recognition systems can be powerful technique to improve recognition performance. The success of these techniques, however, depends on the complementarity of the combined systems. In this paper, a measure of complementarity of different recognition systems is proposed. This measure is based on analysis of similarity of errors made by individual systems. High correlation between the measure and actual performances of combined systems is shown in experiments, which indicates that the measure can be used to select systems suitable for combination. The measure can be computed very efficiently and it can be used even in situations where exhaustive search looking for the set of system optimal for combination would be infeasible.
Combination of different speech recognition systems can be powerful technique to improve recognition performance. The success of these techniques, however, depends on the complementarity of the combined systems. In this paper, a measure of complementarity of different recognition systems is proposed. This measure is based on analysis of similarity of errors made by individual systems. High correlation between the measure and actual performances of combined systems is shown in experiments, which indicates that the measure can be used to select systems suitable for combination. The measure can be computed very efficiently and it can be used even in situations where exhaustive search looking for the set of system optimal for combination would be infeasible.
@inproceedings{BUT17134,
author="Lukáš {Burget}",
title="Measurement of Complementarity of Recognition Systems",
booktitle="Proc. Seventh International conference on Text, Speech and Dialogue",
year="2004",
series="Lecture Notes in Artificial Intelligence (LNAI) subseries of LNCS series as Volume 3206",
pages="283--290",
publisher="Springer Verlag",
address="Brno",
isbn="3-540-23049-1",
url="http://www.fit.vutbr.cz/~burget/publications/burget_tsd2004.pdf"
}