Abstract: Speaker state recognition is an important issue to understand the human behaviour and to achieve more comprehensive speech interactive systems, and therefore has received much attention in recent years. This work addresses the automatic classification of three types of child emotions in vocalisations: neutral mood, fussing (negative mood) and crying (negative mood). Speech, in a broad sense, contains a lot of para-linguistic information that can be revealed by means of different methods for feature extraction and, in this case, these would be useful for mood detection. Here, several set of features are proposed, combined and compared with state-of-art characteristics used for speech-related tasks, and these are based on spectral information, bio-inspired ear model, auditory sparse representations with dictionaries, optimised wavelet coefficients and optimised filter bank for cepstral representation. All the experiments were performed using the Extreme Learning Machines as classifier because it is a state-of-art classifier and to achieve comparable results. The results show that by means of the proposed feature extraction methods it is possible to improve the performance provided by the baseline features. Also, different combinations of the developed feature sets were studied in order to further exploit their properties.
