Candidate selection based on significance testing and its use in normalisation and scoring
Ji‐Hwan Kim, Gil‐Jin Jang, Seong-Jin Yun, Yung Hwan Oh · 1998
Log likelihood ratio normalisation and scoring methods have been studied by many researchers and have improved the performance of speaker identi cation systems. However, these studies have disadvantages: the recognised distorted speech segments are di erent for each speaker. Also the background model in log likelihood ratio normalisation is changed in each speech segment even for the same speaker. This paper presents two techniques. Firstly, candidate selection based on signi cance testing, which designs the background speaker model more accurately. And secondly, the scoring method, which recognises the same distorted speech segments for every speaker. We perform a number of experiments with the SPIDRE database.