Comparative Study on Normalisation in Emotion Recognition from Speech

Abstract

Abstract. The recognition performance of a classifier is affected by various aspects. A huge influence is given by the input data pre-processing. In the current paper we analysed the relation between different normalisation methods for emotionally coloured speech samples deriving general trends to be considered during data pre-processing. From the best of our knowledge, various normalisation approaches are used in the spoken affect recognition community but so far no multi-corpus comparison was conducted. Therefore, well-known methods from literature were compared in a larger study based on nine benchmark corpora, where within each data set a leave-one-speaker-out validation strategy was applied. As normalisation approaches, we investigated standardisation, range normalisation, and centering. These were tested in two possible options: (1) The normalisation parameters were estimated on the whole data set and (2) we obtained the parameters by using emotionally neutral samples only. For classification Support Vector Machines with linear and polynomial kernels as well as Random Forest were used as representatives of classifiers handling input material in different ways. Besides further recommendations we showed that standardisation leads to a significant improvement of the recognition performance. It is also discussed when and how to apply normalisation methods.