Depression Detection using Speech Emotion Recognition: Comparative Analysis of Models
Year: 2026
Type: conference-paper
Abstract: Speech Emotion Recognition (SER) has gained significant attention for its ability to make human-computer interactions more natural and supportive, particularly in applications such as mental health monitoring and intelligent assistant systems. In this work, examine the effectiveness of different deep learning approaches using four widely used datasets: SAVEE, RAVDESS, CREMA-D, and DAIC-WOZ. The models evaluated include Multilayer Perceptron (MLP), Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and hybrid architectures that combine CNN with BiLSTM or BiGRU, along with attention mechanisms to better focus on important features. also explore the impact of using reduced feature sets and introduce a confidence- and risk-based strategy to enhance prediction reliability. The results reveal that model performance differs across datasets, largely due to variations in data complexity. The MLP model achieved the highest accuracy of 95 on the SAVEE dataset, while hybrid models showed stronger performance on more complex datasets like DAIC-WOZ. In contrast, the CREMA-D dataset produced comparatively lower results, indicating its challenging nature. Overall, the study highlights that integrating multiple model techniques with feature selection and attention mechanisms can improve the accuracy and robustness of emotion recognition systems, making them more suitable for real-world use.
Language: English
Topic: Emotion and Mood Recognition
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204129274