Analysing the Impact of LSTM and MFCC on Speech Emotion Recognition Accuracy

Identifying emotional states through the analysis of vocalisations is a difficult task in the realm of human-computer interaction (HCI). Many research methodologies have been employed, and more recent research has advocated the use of deep learning algorithms as viable substitutes for the methodologies that are now employed in SER. The Toronto Emotional Speech Set (TESS) dataset is used in this research work to present a speech emotion detection system based on the Long Short-Term Memory (LSTM) model and Mel Frequency Cepstral Coefficients (MFCC) classification. With the use of voice cues, the suggested system intends to appropriately categorise emotions including anger, disgust, fear, happiness, neutral, and sadness. The LSTM model is trained to categorise the emotions, and the MFCC approach is utilised to extract features from the speech signals. The recognition rate of the trained network was then tested using a set of labelled emotion speech samples. The number of neurons and layers are modified for optimisation based on the MFCC accuracy rate. Outside the preview of earlier studies, this study places more focus on speech-based outcomes than on text-based ones.