SPEECH EMOTION RECOGNITION SYSTEM PERFORMANCE ANALYSIS WITH OPTIMIZED FEATURES USING DIFFERENT CLASSIFICATION ALGORITHMS

Affective computing is becoming increasingly significant in the interaction between humans and machines. Emotion recognition of spoken language is a hot topic in Human Computer Interaction (HCI). Understanding a person's physical and mental state could be greatly aided by learning to identify the emotions conveyed in their speech. There are a number of practical uses for emotion recognition from Speech and it has much interest research domain in recent years. Many of the current options, however, are not yet suitable for use in production environments. The system is divided into three phases: features extraction, features selection/dimensionality reduction and classification. The first step involves extracting a wide variety of features, including prosodic and spectral components, Speech and glottal-waveform signals are used to construct long-term statistics. To tell apart associated emotions is a major challenge for SER systems. These features improve the SER's capacity to distinguish between emotions in speech. Inevitably, this high-dimensional feature vector will contain some repetition. In the second phase, the dimensionality of the feature vectors is reduced through the application of feature selection method such as Auto-Encoder technique described by the authors. Next, several classifiers, including K-Nearest Neighbour (K-NN), Logistic Regression (LR) , Support Vector Machines (SVM), Multi-Layer Perceptrons (MLP) and Convolutional Neural Networks (CNNs) are used to the optimize feature vector in stage three. Two widely-cited datasets serve as the basis for experimental evaluations of method efficacy. The Database for Emotions in Telugu Language (DETL) of a native Telugu Language and Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) of English Language are two such collections. Experimentation is carried out with the different Feature Extraction methods such as MFCC, MFCC+∆MFCC+∆∆MFCC, MFCC+∆MFCC+∆∆MFCC+Pitch+ZCR (41-dimension) and Optimized Features (30-dimension) along with the different Classification Methods. From the results, it is observed that the suggested method has potential for use in speech emotion recognition. Furthermore, when compared to other methods in the context, our approach is clearly superior in terms of classification performance accuracy rate. Here CNN model got highest accuracy with Optimized Features for RAVDESS 95.42% and for DETL 98.37%