Emotion State Detection via Speech using Deep Learning

Oral communication is the most efficient and common way to interact with someone. Due to this, voice signalling has attracted researchers for its potential in computer-human communication. Identifying emotions through speech is complex due to various factors like feature selection and vocal diversity. This study uses the RAVDESS and SAVEE datasets and employs techniques including Multilayer Perceptron, Random Forest, Support Vector Machine, k-Nearest Neighbours, Logistic Regression, Gaussian-Neural Network, and Convolutional Neural Network. The identified features encompass the MelFrequency Cepstral Coefficients (MFCC), Chroma, and Mel spectrogram. To conduct a comparison of accuracy rates, we employed seven classifiers across two datasets. Consequently, CNN model demonstrated superior performance in both the RAVDESS and SAVEE databases, with accuracy rates of $\mathbf{9 6. 7 5 \%}$ and $\mathbf{9 2. 3 6 \%}$ respectively, surpassing other classifiers.