A Music Labeling Model Based on Traditional Chinese Music Characteristics for Emotional Regulation.

The effectiveness of emotion regulation based on traditional Chinese music has been verified in clinical trials over thousands of years, but the reasons are unclear. This paper aims to use feature engineering to find effective music features which are effective for classifying different types of music and thus to try to provide an automatic recognition framework for building music libraries that can be used for mood regulation and music therapy. In this work, five modes (equivalent to the scales of Western music) of traditional Chinese music repertoire which can be used to regulate loneliness, anxiety, anger, joy, and fear are used. Features including Chroma, Mel-spectrogram, Tonnetz, and full feature vector features, are extracted for different length fragments of a piece of music which are then used to build a classification model for the five modes using a convolutional neural network (CNN). The results show that the highest 5-classes classification accuracy, 71.09%, is achieved from a Mel map of 5s music clips. A music mode labeling model is then constructed using a weighted combination of the different individual feature models. This model was then qualitatively evaluated on 13 pieces of music in different musical styles, and the results were reasonable from a music theory perspective. In future work, this music labeling model will be tested on more types of tracks to better assess its reliability.