MS-ACNN: multiscale CNN attention network for music emotion recognition

Music Emotion Recognition (MER) is a subfield of Music Information Retrieval (MIR), widely applied in areas such as music retrieval, music therapy, and music generation. In recent years, MER has gained significant attention in academia. Due to the varying bar lengths across different types of music, this study proposes a neural network model that combines convolutional neural networks (CNNs) with different kernel sizes and a self-attention mechanism. This allows the model to extract local features across varying bar lengths and capture temporal features in transitions between different sections of the music. Additionally, we employed the deep pre-trained L3-Net as the feature extractor for audio data. Our model achieved an accuracy of 76.53% on the TAFFC dataset and 70.83% on the EMOPIA dataset, with AUC scores of 0.9052 and 0.9202, respectively.