MelodyForge: A Context-Adaptive Hierarchical Neural Framework for Automatic Symbolic Music Generation Through Multi-Level Musical Representation Learning

Abstract

The production of coherent and meaningful symbolic music, in a musical sense, is a challenging task for AI. The existing approach makes it possible to make the sequence of notes reasonably sound, but does not do a good job of keeping the continuity of a musical composition and the surrounding context. We propose a context-adaptive hierarchical neural network called MelodyForge that learns multiple different levels of musical patterns, and then we discuss the use of this network in the enhancement of symbolic music generation. The proposed framework is built around a Context-Adaptive Hierarchical Encoder and the Multi-Level Musical Representation Learning module and on a Hierarchical Music Decoder that models the context to generate coherent musical sequences. The effectiveness of the developed framework was evaluated using Accuracy, Precision, Recall, F1-score and BLEU score as well as Perplexity and was compared with some of the existing symbolic music generation models. The proposed model had an accuracy of 91.20%, BLEU score of 0.487 and Perplexity of 2.81, which proved to be consistent with comparison methods. Qualitative assessment also found that the music produced is more melodic, has a better balance of the rhythm patterns and has more continuity in the structure. Furthermore, an ablation study was performed to confirm the contributions of each component of the architecture proposed to obtain the final performance. The results show that the symbolic music generation framework of MelodyForge has proved to be effective, and it will be used to conduct further research in symbolic music generation with expressive and styleaware approach and multi-instrument music generation.