Optimizing Speech Emotion Recognition with Dynamic Dilation Rates for Efficient Edge Deployment

Speech emotion recognition (SER) has broad applications, from aiding individuals who struggle to express emotions to supporting mental health assessments. Detecting negative emotions, such as anger and sadness, is particularly crucial for long-term emotional monitoring. However, achieving reliable SER on resource-constrained edge devices presents challenges in computational efficiency and data scarcity. This research proposes an efficient SER model optimized for edge deployment by leveraging a temporal convolutional network (TCN) with dynamically adjustable dilation rates, reducing computational complexity while maintaining high recognition accuracy. Additionally, transfer learning is employed to mitigate data scarcity, enabling robust recognition across different users and datasets. Experiments were conducted in two scenarios: cross-user adaptation within the same dataset and cross-database transfer using the Emo-DB and CHVE corpora. The best results achieved an average recognition accuracy of 82.22% in cross-user adaptation, with a sixfold reduction in model parameters compared to conventional approaches. These results demonstrate that our optimized TCN-based SER model provides an effective and computationally efficient solution suitable for edge computing environments. Furthermore, the developed application successfully enables real-time spoken emotion recognition, paving the way for practical edge AI deployment.