the analysis of the fcm and wknn algorithms performance for the emotional corpus srol

The purpose of this research is to find a set of relevant parameters for the emotion recognition. In this study we used the recordings from the emotion database SROL which is part of the project "Voiced Sounds of Romanian Language". The database was validated by human listeners. The recognition accuracy of the correct expressed emotion (neutral tone, joy, fury and sadness) for the entire database was 63.97%. We used for the classification of input data the Recurrent Fuzzy C-Means (FCM) and WKNN algorithms. We compared the cluster position with the statistical parameters extracted from vowels in order to establish the relevance of each parameter in the recognition of the emotions. For the extracted parameters for each vowel (mean, median and standard deviation of fundamental frequency - F0 and F1-F4 formants, jitter, and shimmer) the FCM algorithm gave satisfactory results in the phonemes recognition, but not to the emotions. For this reason we used WKNN algorithm in classification, which provided the errors around 20-30% comparing with FCM algorithm when the classification errors are around 40-50%.