SVM-BASED EMOTION RECOGNITION FROM SPEECH WITH GTCC AND FREQUENCY FEATURES

Abstract

When a person is in a certain emotional state, a large number of physiological changes occur in the body. These changes significantly affect the way words are pronounced compared to neutral speech. This means that the configuration of the vocal tract changes depending on the speaker’s emotional state. Furthermore, in emotional speech, physiological changes influence certain speech properties, such as speech rate, intensity, and pitchSuccessful classification of emotional speech into the appropriate emotion class requires extraction of salient speech features and construction of a feature vector composed of discriminative attributes that facilitate accurate classification. In this study, we use Gammatone Cepstral Coefficients (GTCC) as components of the feature vector for speech emotion recognition. GTCC are a biologically inspired modification of Mel-Frequency Cepstral Coefficients (MFCC). They are based on gammatone filters, which simulate the human auditory system more effectively than the mel-frequency filters used in MFCCThe remainder of the feature vector is composed of spectral characteristics extracted from the speech signal. In our classification model, the components of the feature vector are primarily extracted by performing spectral analysis on short-time frames of the observed speech signal. Feature vectors constitute discriminative representations that facilitate the more effective classification of speech into corresponding emotional categories. Our classifier is based on Support Vector Machines (SVM), with optimized hyper-parameters.