| Abstract |
Speech recognition is an important research topic that helps understand human emotion and the semantics associated with it. The emotional state of a person is recognized using speech emotion recognition systems. These systems do not take into account the semantic aspects of the person. This task of understanding human emotions using programmed architectures like applications or websites is an active part of ongoing research. It aids in a wide array of applications ranging from human-computer interface models (HCIs) to medical diagnosis of mental health. Most of the existing research fails to classify emotions for the given dataset efficiently; moreover, the feature extraction methodology is absent in related work. In this chapter, we aim to study the linear and non linear features of human speech and its classification into different emotions like anger, sad, pleasant surprise, happy, and relaxed state. By properly distinguishing different human emotions using emotion classifiers, we aim to develop smart cities using efficient and real-time methods of emotion detection through cell phone conversations, call center applications, cabdrivers, and other machine-dependent communication sources. We realize that our proposed architecture is mainly useful in environments that are assisted by interaction devices, but our work can potentially also be applied in the medical field where the treatments or reactions to medicine can be studied. We aim to analyze different classifiers on our dataset by using classifiers such as SVM, Random Forest, RNN, MLP, Gradient Boosting Classifier, KNN, Bagging Classifier, etc. based on different algorithmic metrics. The dataset is a mixture of two different databases containing recordings from multiple actors and actresses. Filtering and analyses of different classifiers along with their confidence matrix are studied. During experimentation, the MLP classifier considerably outperforms all the other classifiers with its accuracy accompanied by neural networks that also show promising results. © Ron Waddams/Bridgeman Images. |
