End-to-end music emotion variation detection using iteratively reconstructed deep features

Automatic music emotion recognition (MER) has received increased attention in areas of music information retrieval and user interface development. Music emotion variation detection (or dynamic MER) captures also temporal changes of emotion, and emotional content in music is expressed as a series of valence-arousal predictions. One of the issues in MER is extraction of emotional characteristics from audio signal. We propose a deep neural network based solution for mining music emotion-related salient features directly from raw audio waveform. The proposed architecture is based on stacking one-dimensional convolution layer, autoencoder-based layer with iterative reconstruction, and bidirectional gated recurrent unit. The tests on the DEAM dataset have shown that the proposed solution, in comparison with other state-of-the-art systems, can bring a significant improvement of the regression accuracy, notably for the valence dimension. It is shown that the proposed iterative reconstruction layer is able to enhance the discriminative properties of the features and further increase regression accuracy.