Music Emotion Classification with Source Separation Based MSB-Conformer

The paper presents a novel approach for classifying music emotions that addresses several limitations, including limited emotion categories, low accuracy, and counterintuitive results compared to the Valence-Arousal (V-A) model. The proposed method involves separating music sources using the MULTI-SCALE MULTI-BAND DENSENETS model, simplifying the complex music structure into vocals, drums, bass, and four other audio tracks. Time-series features and fixed attribute features are extracted from each track. The time-series features are inputted into the Conformer model, which utilizes convolutional neural networks for local feature extraction and Transformer encoders for capturing long-distance dependencies. The Conformer model output is combined with fixed attribute features and passed through fully connected layers for emotion classification. Experiments on a Netease Cloud Music dataset with 12 emotion categories demonstrated that the proposed MSB-Conformer model outperformed comparative models (CNN+LSTM, WaveNet, Transformer), achieving an average accuracy of 94.24% and surpassing the state-of-the-art in classification categories. The effectiveness and generality of the model were validated using the Emotify dataset. The proposed method simplifies music complexity through source separation and utilizes the Conformer model to simultaneously model local and global audio features, offering a novel approach to music emotion classification tasks.