A New Framework Toward Accurate Emotions Recognition for Human Speeches

The well-known Multi-Layer Perceptron (MLP) classifier is proposed in a new manner to recognize emotions of the speaker by analyzing her/his speech. The CREMA-D emotional dataset is used for training and evaluating the developed recognition system. Six emotion” s types are examined: namely, sadness, anger, neutral, disgust, happiness, and fear. A separate MLP classifier is trained for each emotion” s type, considering all other emotions as “unknown”. The framework is four phases. 1) Data preprocessing, splitting, and normalization, 2) feature extraction and/or selection, 3) model creation and fitting, 4) model evaluation. Several experiments were conducted to explore the efficiency of the developed framework. Results of these experiments show that the new framework is efficient. The accuracy achieved is 83% that outperforms all the recent works in the literature for the same datasets.