Multi-features learning via Attention-BLSTM for Speech Emotion Recognition

Within the domain of human-computer interaction, Speech Emotion Recognition (SER) has assumed critical importance, where the careful selection of feature representation stands as a key determinant. This study is driven by the objective of conducting a comprehensive comparative analysis through the implementation of a parallel multi-feature design, strategically aimed to effectively characterize the emotional information within speech signals. The innovation of our proposed approach lies in the integration of frameworks, including ImageBind, WavLM, HuBERT, Wav2vec 2-xlsr, along with the commonly used MFCC features. Leveraging their knowledge and adaptability, we tailor these frameworks to meet the specific demands of SER task. Furthermore, to encompass a rich set of multi-features, we introduced an Attention-based Bidirectional Long ShortTerm Memory (attention-BLSTM) network to capture and learn the distinct emotional speech representations extracted through various methods. The experimental evaluations were executed using the IEMOCAP database, and the most remarkable results were attained by employing ImageBind embeddings with our proposed attention-BLSTM, achieving an accuracy of $81.03 \%$. This outcome outshines the current state-of-the-art approaches in SER.