부산대학교 도서관

This paper proposes a CBGRU-based speech emotion recognition model with a multi-headed self-attentive mechanism to address issues such as long training times, insufficient correlation of feature information, and low recognition rates in speech emotion recognition research. The BGRU network extracts short-term and long-term data correlations to enhance data utilization, while the introduced multi-head self-attention mechanism quickly computes input data sequence and re-allocates weights at different positions to prevent over-focusing on one’s own position, thus making the extracted features more representative. Sentiment classification is performed accurately using the softmax function. Experimental results on two corpora demonstrate an average recognition rate of 71.46% on a self-built electricity customer service corpus. Comparison of multiple experiments shows that BGRU reduces model training time, while the embedding of the multi-headed self-attentive mechanism improves recognition performance.