الفهرس 
INTRODUCTION AND MOTIVATIONSOnly 14 pages are availabe for public view
 |  | from | 135 |  |  |
| | | from | 135 | | |
Abstract
In past years, different researches have paid attention to the recognition of nonverbal information, and have especially focused on emotion recognition. Many kinds of physiological characteristics are used to extract emotions, such as voice, facial expressions, hand gestures, body movements, heartbeat and blood pressure. Among these modalities, facial expressions and speech are known to be more effective, for expressing the emotions.
Speech Emotion Recognition (SER) can find several applications such as call centers management, commercial products, life-support systems, virtual guides, customer service, lie detectors, conference room research, emotional speech synthesis, art, entertainment and others.
This work emphasizes on recognizing different emotions from speech signals. Working on a simulated type of speech data base, the classification system was introduced in four main steps each was clearly explained. The classification process starts by extracting some emotion-related ”features” from the human speech. The extracted features are related to statistics of pitch, formants, and energy contours, as well as spectral, perceptual and temporal features, jitter, and shimmer. These extracted features are then processed and used by the classifier to guess the speaker’s emotion. In this work, Artificial Neural Networks (ANN) was chosen as the classifier. Working on finding a robust and fast ANN classifier suitable for different real life application was our target.
Using five different ANN architectures, a large on line database with 175 extracted features from different speeches, it will be shown that different factors have high impact on the classification success rate, namely, the data format, training data combinations, number of features and the ANN architecture and design. Several experiments were carried out to show how these factors can be tuned to achieve a success rate of 85% and even more. It will also be shown that the proper tuning of these factors can help reducing the number of features required for classification, and hence achieve a computationally efficient implementation for the classifier, opening a wide field of different online, low cost applications.