Fusion of temporal-spectral features and transformer architectures for automated music mood classification

Music profoundly influences human emotions, mental health, and well-being, making accurate mood recognition crucial for applications such as personalized recommendations and therapeutic interventions. Traditional music mood classification methods rely on handcrafted audio features and classical machine learning algorithms, which often struggle to capture complex temporal and spectral patterns in music. With advances in artificial intelligence, transformer-based models have emerged as powerful tools for analyzing global contextual information in sequential data, offering new opportunities for improved mood classification. In this study, we propose the use of the Patchout Spectrogram Transformer (PaSST) to classify music moods. Our approach integrates audio-based features including waveform patterns, Mel-spectrograms, chromagrams, and Root Mean Square (RMS) energy, combined with an attention mechanism that captures intricate acoustic relationships. We perform empirical evaluations comparing PaSST with traditional convolutional on a publicly available dataset of approximately 2,000 music tracks. The results demonstrate that the proposed PaSST model significantly outperforms existing models, achieving a highest classification accuracy of 95%, which highlights the effectiveness of transformer architectures in music mood recognition beyond traditional method.