Anchor Models and WCCN Normalization For Speaker Trait Classification

摘要

This paper presents an improved version of anchor model applied to solve the two-class classification tasks of the INTERSPEECH 2012 speaker trait Challenge. To build the anchor model space of each task, we include the class models of all tasks. The introduction of within-class covariance normalization (WCCN) applied to the log-likelihood scores of the anchor space not only improves the results compared to the unnormalized version but also exceeds the performance of GMM or GMM-UBM systems. Even if Euclidean distance gives worst performances compared to cosine metric, we find that after normalization both metrics give similar results so they can be used interchangeably.