Yuan Li
Bayesian, Transformer, GAN, music generation, confidence, posterior probability
Music is rich in elements such as emotions, rhythms, and harmonies, and there are significant differences among different styles. As long-sequence data, music has complex dependency relationships. This complexity makes it impossible to accurately generate music that conforms to style rules and evaluate its generation confidence. Therefore, a method for evaluating the confidence of music generation based on Bayesian Transformer is proposed. The Transformer–generative adversarial network (GAN) network combines the advantages of Transformer and GAN to achieve the generation and evaluation of multi-track music. In this network, the generator first generates single-track music for a single instrument and then inputs it into the Transformer module. The Transformer generates single-track music containing the track information of the target instrument by learning the track features of two instruments. The single-track music of different instruments is combined to form multi-track music, and the effect is optimised through the discriminator and music rules. The generated multi-track music constructs an auto-regressive model in the form of a time series. By calculating the residual sequence between the predicted value and the real value and combining the Bayesian method to analyse its prior probability and conditional probability, the posterior probability of the time series being normal or abnormal is finally obtained. Based on the logarithmic ratio of the normal and abnormal posterior probabilities, the system determines whether there are abnormalities in the multi-track music generation time series. If an abnormal value is detected, the time series is input into the self-organising map neural network (SOM), and the music generation confidence is calculated using the trained state transition matrix and the current state sequence, thus completing the quantitative evaluation of the music generation quality. The experimental results show that: in the proposed method, the multi-track music generation process of the Transformer–GAN network containing three instruments, namely piano, guitar, and bass, is from overall to details, and the generated music quality is extremely high; moreover, after inputting the abnormal multi-track music into the SOM, the confidence of the abnormal multi-track music is only 31%, which is extremely low compared with the confidence of 97% for normal music. ∗ Department of Music, College of Art, Shandong University, Weihai, Shandong, 264209 China; e-mail: [email protected] Corresponding author: Yuan Li Recommended by Jian Su
Important Links:
Go Back
