Abstract

The focus of this thesis is on emotion recognition based on the speech signal. The state of the art is being reviewed. Based on models in psychology and the requirements of automatic systems, models for emotion recognition from speech are proposed and a most appropriate one for automatic detection is chosen.
The aim is to search for methods suitable for enhancing the generality, portability and robustness of emotion recognition systems. For this purpose we introduce a minimal set of features that provides recognition capabilities. The set has been tested on multiple databases and the results show that it provides capabilities for emotion discrimination.
An experiment including more databases was designed, aiming to get insight into the generalization capabilities of systems trained on extended corpora, as well as their portability. The results did however not give a clear indication in that sense.
Several classifcation techniques along with dierent feature types were fused together and better performing systems were generated. The best results were determined by the use of logistic regression fusion on t-normalized data.
Besides a list of experiments on acted corpora, two databases containing real speech have been used. The framework generated by real data was dierent than for acted data. The problem of working with emotions in a continuous space was also touched.
The knowledge gathered for all the experiments was used for building EmoReSp, a real-time system for emotion recognition based on speech.