Emotion Recognition From Speech Using Telugu Language and Deep Learning Models

This research involves the use of deep learning techniques to detect emotions, in Telugu speech signals, a research domain that has received little attention. Amrita speech emotion dataset-1 (ASED-1) consisting of 550 signals representing five categories. Surprise, anger, happiness, neutrality, and sadness. We explore how well various deep learning models can recognize emotions from speech. Our approach involves training models such as Convolutional Neural Networks (CNNs) Long Short-Term Memory (LSTM) networks and Multi-Layer Perceptron (MLPs) on a portion of the data and evaluating their performance on a test set. The MLP model achieved an accuracy rate of 84% although this was constrained by the size of the dataset. The CNN model showed results with decreasing trends in both training and validation loss indicating learning and generalization abilities. Similarly, the AlexNet architecture demonstrated a decline in both training and validation loss suggesting a fit without overfitting. Additionally, we included a machine learning classifier at the layer. Observed different accuracies across various models. The RNN model achieved an encouraging accuracy rate on the test data showing potential for emotion classification tasks but also room for improvement, through complex architectures or larger datasets. This research paper adds to the knowledge, in the field of emotion recognition, in Telugu speech by emphasizing the difficulties and possibilities of using deep learning techniques in this domain.