Speech Emotion Recognition (SER) has been one of the shining fields of deep learning related human-computer interaction technology to analyze human emotions through speech signals. The aim of this paper is to recognize multiple emotional states from the audio signals taken from the RAVDESS dataset. A CNN model is employed in this work because the CNN can learn the local features well from two-dimensional representations of speech signals such as Mel Spectrogram. Traditional machine learning methods suffer from low accuracy when compared with contemporary deep learning solutions in emotion classification. The proposed system accepts audio files as input, then performs preprocessing and feature extraction with Mel Spectrograms. In order to enhance model robustness, data augmentation approaches are also employed. The CNN is then trained on the extracted features for accurate prediction of emotion. Experiments demonstrate that the proposed method attains a classification accuracy of around 95%, proving its effectiveness in detecting certain emotions in speech data.
Nithisha et al. (Wed,) studied this question.