This paper presents an improved version of anchor model applied to solve the two-class classification tasks of the INTERSPEECH 2012 speaker trait Challenge. The introduction of within-class covariance normalization applied to the log-likelihood scores of the anchor space can not only improve the results compared to the unnormalized version but also exceed the performance of GMM or GMM-UBM systems. Furthermore, our results on development set show a relative improvement of 6.1%, 8.6% and 3.2% for the Personality, likability and pathology sub-challenges respectively compared to the best baseline systems provided by the organizers.

Index Terms: anchor model, WCCN, speaker trait classification, GMM model, Interspeech 2012 challenge