Semi-Supervised Self-Learning Enhanced Music Emotion Recognition

The music emotion recognition (MER) task aims to recognize the emotion expressed in a given music clip automatically. Music emotion recognition can be widely used in many fields, such as dynamically generating music to adapt to the emotion of scenes in movies or games [1], music-assisted psychological or physical therapy, personalized recommendation in stream media, human-machine interaction, music retrieval, and so on, which has broad application prospects. In recent years, As the amount of data grows, data-driven deep learning methods have become the mainstream method in the Music Information Retrieval (MIR) field [2, 3]. At present, the duration of audio clips in public music emotion datasets is 30 45 seconds. Although the longer the duration is, the more helpful it is to distinguish emotions, according to the study of music psychology, it is found that the duration of about one second of music is sufficient to evoke an emotional reaction [4]. To address the issue of limited annotated data in emotion recognition tasks, some segment-based methods [5-8] have been proposed recently, which naturally increase the amount of training data and can make full use of every audio sample in the dataset.