NSER: Development of Novel Speech based Emotion Recognition System by using Hybrid Learning Technique

For a long time, emotion identification from speech signals has served as an important area of study for HMI applications. Emotion detection systems have already been created for a wide variety of signals. The training and evaluation subsets were manually created using the open-source RAVDESS dataset, which contains acted voice recordings from both male and female speakers of English. Mel-frequency cepstral coefficients (MFCCs) were employed to represent vocal tract information of the speakers. These coefficients were extracted from audio samples within the training dataset. Additionally, feature extraction was carried out on recorded natural speech. The resulting feature vectors were directed to a Novel Hybrid Learning Technique. The initial processing stage of this technique involves a Convolutional Neural Network (CNN), which facilitates the extraction and transformation of audio features through convolutional layers. These layers are intentionally designed to capture pertinent patterns and features in audio signals, a crucial aspect for comprehending emotions. Incorporating Support Vector Machine (SVM) algorithm, known for its efficiency in handling high-dimensional feature spaces, enhances the system's performance in tasks like audio emotion detection. As a result, the proposed model stands out with an outstanding accuracy rate of 96.75%, underscoring its remarkable capability for precise emotion recognition.