8th International Conference on Spoken Language Processing

Jeju Island, Korea
October 4-8, 2004

Improvement of Confidence Measure Performance using Background Model Set Algorithm

Byoung-Don Kim (1), Jin-Young Kim (1), Seung-Ho Choi (2), Young-Bum Lee (3), Kyoung-Rok Lee (4)

(1) Chonnam National University, Korea
(2) Dongshin University, Korea
(3) Samsung Advanced Institute of Technology, Korea
(4) Nambu University, Korea

This study suggests the BMS (Background Model Set) algorithm used in the speaker verification to supplement the shortcoming in the process of calculation of RLJ-CM (RLJ-Confidence Measure) and normalized CM. The confidence measure shows the relative similarity between the recognized model and the unrecognized one. In calculation of the CM, the composition of anti-phone model does not have the high confidence measure because a probability and a standard deviation are calculated by using all the phonemes. Also, there is shortcoming that the recognition time increases at the calculation using all phonemes. To solve such a problem, the method is researched which re-organizes a probability and a standard deviation by using the BMS algorithm. As a result, when the BMS algorithm was applied near 17% of the MDR(Missed Detection Rate), a performance was increased to 0.104FA/KW/HR(false alarm/keyword/hour), by 50% improvement compared to the model not applied with the BMS. While the existing work recognition took on average 15 minutes to deal with the evaluation database of one person, the recognition execution through the BMS base reduced the recognition time to 10 minutes by 33%.

Full Paper

Bibliographic reference.  Kim, Byoung-Don / Kim, Jin-Young / Choi, Seung-Ho / Lee, Young-Bum / Lee, Kyoung-Rok (2004): "Improvement of confidence measure performance using background model set algorithm", In INTERSPEECH-2004, 2113-2116.