Evaluation of emotion recognition from speech

Over the last few years, interest on paralinguistic information classification has grown considerably. However, in comparison to related speech processing tasks such as Automatic Speech and Speaker Recognition, practically no standardised corpora and test-conditions exist to compare performances under exactly the same conditions. The successive challenges proposed at the world´s largest conference on automatic speech processing, namely the INTERSPEECH conferences, are important for comparing performance of statistical classifiers. In this paper, we summarize results, commonly used methods of challenge participants and results of Koç University, Multimedia, Vision and Graphics Laboratory on the same tasks. Our main contributions include Formant Position-based weighted Spectral features that emphasize emotion in speech and RANSAC-based (Random Sampling Consensus) Training data selection for pruning possible outliers in the training set.