The Sound of Sentiment: Optimizing Emotion Recognition in Speech with Neural Networks

Speech emotion recognition is the identification of a speaker's emotional state from speech signals, which becomes very critical in human-computer interaction. This study aims at showing how advanced machine-learning techniques can be used to increase the accuracy of emotion recognition through audio data. Among the key features used in EDA are MFCC and Mel spectrograms, which are very efficient at capturing every subtlety of human speech. The models to be compared for emotion classification include decision trees, K-nearest neighbors, multilayer perceptron, gated recurrent units, long short-term memory, and convolutional neural networks. Decision Tree Overfitting: The result on the train set was 64.8%, and on the test set, its average score was 51.0%. The KNN model performed moderately well; it scored 95.0% on the training set and 67%on the test set. Of course, the MLP classifier did very well on the train set but failed dramatically when generalizing the test set with an accuracy of 23.8%. The test accuracy of the GRU model was 67.3%. The LSTM model had an accuracy equal to 85.0% in the test, with a very high effectiveness against a data sequence. Of all these, the CNN model did better and achieved an overall test accuracy of 85.0%. The classification report contains precisions, recalls, and F1 scores for several emotional categories, which prove the strength of the CNN model. This paper thus puts forward the potential of using CNNs in SER applications and provides a general performance analysis of the proposed model.