Speech Emotion Recognition Using Weighted Score Fusion for Low Resource Consumer Devices

Deep learning techniques have significantly advanced machine performance in tasks traditionally dominated by human expertise. One such task is speech emotion recognition (SER), a crucial component of affective computing. Given the inherent complexity of recognizing emotions, even for humans, there is a persistent demand for models that are both robust and accurate, yet less complex for deployment on low-resource devices. Despite notable progress in SER systems, achieving high performance with low-parameter models remains a challenge. In this paper, we introduce a transformer-based multi-layer SER model (TBWSC model) that employs a weight score configuration (WSC) across shallow, intermediate, and high-level learned features. This approach enhances SER performance by utilizing a transformer encoder to capture detailed acoustic feature representations, reducing the need for a deep network. Our model achieves significant improvements with just three layers, compared to traditional deep architectures. Additionally, we performed a comprehensive evaluation of our proposed TBWSC SER model to highlight the benefits of weighted fusion and complexity reduction. The model was tested on multiple datasets, including the emotional German speech dataset (EMODB), the Surrey audio-visual expressed emotion (SAVEE) database, and the Toronto emotional speech set (TESS). These datasets encompass emotional states such as happiness, sadness, neutrality, and anger. Our TBWSC model demonstrated superior performance, achieving an average accuracy of 81.82% on EMODB, 82.22% on SAVEE, and 99.65% on TESS. These results underscore the effectiveness of our weighted feature fusion approach and the potential for deploying efficient SER systems in low resource consumer devices.