Predicting Likability of Speakers with Gaussian Processes

In this paper, we study the problem of predicting likability of speakers based on their voices. We use the data provided by the Likability Sub-Challenge of Inter Speech 2012 Speaker Trait Challenge. We explore Gaussian Processes (GP) based learning techniques for classification and regression. In particular, we propose a novel algorithm that greedily finds a sparse yet informative subset of features. We also show how the covariance functions learned for GP models can be used to derive new features for prediction. The best system that we have developed integrates those techniques and achieves 60.1% unweighted accuracy on the Likability Sub-Challenge (1.86% relative improvement over the provided baseline). The proposed approach also works well on the Pathology Sub-Challenge, achieving an accuracy of 73.7% ( 6.92% relative improvement).