Speech Emotion Recognition: Leveraging LSTM and Change Point Detection for Continuous Emotion Detection

Emotion recognition in continuous speech remains a significant challenge for human-computer interaction, as current approaches focus on discrete segments, missing the continuous flow and transitions of emotions in natural conversations. This research proposes a dual-stage system combining deep learning and change point detection to identify discrete emotions and their transition points in continuous speech, addressing the limitations of current approaches. The system employs long short-term memory networks for emotion classification. It is trained on the Interactive Emotional Dyadic Motion Capture database, achieving an accuracy of 91.04% in identifying discrete emotional states. We then implement the Pruned Exact Linear Time algorithm to detect points where emotional transitions occur in continuous speech. The detection accuracy ranges from 19.40% at quarter-second tolerance to 89.43% at two-second tolerance. These findings represent a significant advancement in continuous emotion recognition technology, with practical applications in mental health monitoring, customer service analysis, and human-computer interaction systems.