On the use of statistical ensemble methods for telephone-line speaker identification

To improve the performance of the telephone-line speaker identification system, we use three statistical ensemble methods: boosting, bagging and a combination of weak perceptrons (CWP). Our motivation lies in two aspects. Firstly, we try to find out whether those statistical ensemble methods can work when they are applied on this given real world problem, although they perform well in several other learning problems. Secondly, unlike previous researches using statistical ensemble methods where no studies are conducted on mismatch conditions, we investigate our system where miscellaneous mismatch factors are involved. We also investigate the stability performance of three statistical ensemble methods that reflects their capability against mismatch.