Emotion recognition from children's speech using anchor models

In this paper we have adopted anchor models to solve a multi-class problem of automatic emotion recognition from children's speech. The likelihood scores of an utterance over the emotion models are normalized using their within-class covariance matrix (WCCN) in order to increase the difference in the characteristic behavior of scores between classes. After normalization, we find that WCCN not only increases performance but also produces similar performances for cosine and Euclidean metrics. We also show that, in contrast to speaker diarization and verification problems, the performance of anchor model exceeds GMM’s performance by a relative gain of 6.2%. Finally, anchor model improves the state-of-the art by 2.6% relative.