MFAN: Multi-Scale Feature Attention Network for Speech Emotion Recognition

Due to the variability of speech signals and the complexity of human emotions, speech emotion recognition (SER) is an important and challenging task. It is crucial in SER to extract rich emotional information. Some research relies on one type of feature to extract emotional information. However, this approach often fails to capture rich emotional information. Some studies employ multiple types of features. Different features enhance the richness of emotional information, but it may also be affected by redundant information, leading to a decrease in recognition performance. In this paper, we propose a multi-scale feature attention network (MFAN) to address these problems. The model utilizes an utterance-level feature extraction module (UFEM) and a multi-scale feature extraction module (MFEM) to extract emotional information at different scales. In UFEM, a global perspective is being used to extract utterance-level emotional features. In MFEM, a multi-scale feature attention mechanism is introduced to extract emotional information at various scales from a local perspective. The attention mechanism is to enable the model to focus on information relevant to emotion recognition, reducing the impact of redundant information and improving the overall performance of the model. The experimental results on the IEMOCAP, RAVDESS, and EMODB datasets demonstrate that MFAN exhibits excellent performance in the field of speech emotion recognition.