Abstract
This paper presents the results of cross-corpus emotion recognition in the context of Arabic and English languages. Two corpora are used: the King Saud University Emotions (KSUEmotions) corpus for Arabic speech and the Emotional Prosody Speech and Transcripts (EPST) corpus for English speech. Mapping emotions in speech signals into valence (positive and negative) and arousal (low and high) are performed on each corpus as well as on a mix of the two corpora. Low-level acoustic features are extracted, and the Deep Belief Networks (DBN) and Multi-Layer Perceptron (MLP) classifiers are used to perform experiments in this study. The results obtained are shown to be better when the corpora are divided into arousal classes, than when the corpora are divided into valence classes. The DBN classifier achieves an emotion recognition rate of 69.91% on the mixed corpus. In general, the DBN classifier is better than the MLP classifier; however, when we use only two classes, the performance results are nearly equal.