Exploring the EmoBone Dataset with Bi-Directional LSTM and Attention for Emotion Recognition via Bone Conducted Speech

This research investigation explores the use of deep learning for emotion recognition using bone-conducted (BC) speech. It focuses on the EmoBone dataset, which is the first dataset specially created for this purpose. Bone-conducted speech offers several advantages, such as improved speech quality in noisy environments, reduced interference from ambient sounds, and enhanced user privacy, as the sound is directly transmitted to the auditory system through bone conduction. Although deep learning is commonly employed for voice emotion recognition, these methods frequently encounter challenges, including the degradation problem and information loss in the deeper stages of deep neural networks. To resolve these issues, we propose implementing a bi-directional long short-term memory (BiLSTM) network that utilizes BC speech and incorporates attention mechanisms. This network will effectively utilize information and mitigate the adverse impacts caused by degradation. However, the attention mechanism significantly improved the performance of the model, resulting in an accuracy gain of approximately 91.45%. The confusion matrices demonstrate that the attention mechanism enhances recognition accuracy for all emotional categories. This study exhibits the capability of enhancing emotion recognition in BC speech applications by integrating BiLSTM with attention.