The Music Generation Road from Statistical Method to Deep Learning

Abstract

Automatic music generation is regarded as one of the most concerning research areas. However, the current music generation AI model lacks mainstream algorithms and norms, including model evaluation standards and objective music quality measurements. In this paper, we conducted experiments on traditional statistical models, machine learning models and deep learning models to generate music, respectively involving Markov Chain, SVM, Random Forest, LSTM, GAN models and their variants. In addition, we also proposed three new quantitative methods to measure the quality of computer-generated music using technical metrics: the Key Confidence Test, the Random Walk Test and the Word2Vec Similarity Test. These methods could be used to measure the quality of generated music and avoid the subjectivity of manual evaluation, thus providing reference for future music evaluation tasks.