Speech Emotion Recognition Using Neural Network and Wavelet Features

Human speech which is generated through the vibration of the vocal cord gets affected by the emotional state of the speaker. Accurate recognition of different emotions concealed in human speech is a significant factor toward further improvement of the quality of Human–Computer Interaction (HCI). But the satisfactory level of accuracy is not yet achieved mainly because there is no well-accepted standard feature set. Emotions are hard to distinguish from speech even by human and that is why the standard feature set is difficult to extract. This paper presents a model to classify emotions from speech signals with high accuracy compared to the present state of the art. The speech dataset used in this experiment where speech recordings that are specifically labeled with different emotions of the speakers. A wavelet-based novel feature set is extracted from speech signals and then a Neural Network (NN) with a single hidden layer is trained on the feature set for classification of different emotions. The feature set is a newly introduced one and for the first time it is being tested with NN architecture and classification results are also compared with the results of other prominent classification techniques.