Published Paper Details:

This study explores emotion recognition in speech using Long Short-Term Memory (LSTM) networks on the Toronto Emotional Speech Set (TESS) dataset. Mel-frequency cepstral coefficients (MFCC) features are extracted, and an LSTM model is trained with a specific architecture, including Rectified Linear Unit (ReLU) and Linear activation functions in dense layers to get the membership of an audio sample in all the 7 labels and also compare them.The compiled model, utilizing mean squared error as the loss function and the Adam optimizer, achieves a training accuracy of 99% and a validation accuracy of 100%. Notably, our approach includes feature membership analysis, providing insights into the contribution of individual features to each emotion label.This comprehensive analysis demonstrates the efficacy of the proposed LSTM architecture in accurately identifying emotional states in speech, while the feature membership analysis adds a nuanced layer of interpretability to the model's predictions. The success of our approach suggests potential applications in emotion-aware technologies and interactive systems.