Analysis of AI Evaluation for Music Pieces

In recent years, an upsurge of Artificial Intelligence models and other machine learning methods has renovated the field of computer music, especially rating music using computers. A lot of methodologies are only based on numerical features of music to predict popularity scores, and the input parameters are often with limited information about the actual music. Moreover, they often focus on already existed features for published music. This research, on the other hand, introduces an improved method, which predicts the popularity score of raw audio file. The methodology includes a series of pre-trained audio feature extraction models, YAMNet and Librosa, to get audio feature data. The research also includes a RandomForest regression model that uses both numerical and categorical features to predict the popularity score of the input music. Tested results have shown a mean squared error (MSE) of 183.542 and a percentage error of 16.89%. This research emphasizes the possibility to advance music performance rating models as well as leading the direction for more well-developed advanced music rating applications.