Speech Emotion Recognition (SER) is an evolving field aimed at enhancing human-computer interaction by enabling machines to understand human emotions from voice. This study presents a deep learning-based SER system using Long Short-Term Memory (LSTM) networks trained on the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS). The input speech signals are processed to extract Mel Frequency Cepstral Coefficients (MFCCs), which are fed into a sequential LSTM model to classify emotions such as happiness, sadness, anger, and calmness. Our model was trained and evaluated using an 80–20 train-test split, and it achieved a training accuracy of 95% and a validation accuracy of approximately 82.3%. These results are promising for building more emotionally aware interfaces in applications like virtual assistants, customer support bots, and healthcare systems. In this paper, we also examine how LSTM networks handle time-series audio data effectively, discuss the performance compared with existing literature, and analyze potential areas for further development. Our findings confirm that LSTM with MFCC is a viable approach for SER tasks and lays the groundwork for more advanced emotion-aware technologies.