Music emotion analysis using semantic embedding recurrent neural networks

The paper presents an original approach to music emotion recognition. We propose to use recurrent neural networks to separate the representation learning process from the classifier, which allows us to use a Support Vector Machine on top of a network to improve the results. We define a suitable loss function that is able to find a feature space in which similarity between vectors representing the music recordings corresponds to the similarity between their annotations. The proposed method was tested for regression and classification using two datasets. The results are presented and discussed.