Harmonizing AI-Generated Music: Integrating Symbolic and Audio Models for Text-to-Music Generation

The evolution of AI-generated music through input text has seen remarkable advancements in both symbolic and audio music generation. Despite this progress, the synergy between these two domains remains underexplored. Consequently, we introduce a novel method for text-to-music generation, capitalizing on the precise control over specific musical attributes provided by symbolic music models and the ability of audio music models to generate music coherent with the contextual meaning of input text. This method enhances the alignment between the generated music and the input text. Specifically, the proposed method initiates by generating symbolic music from input text, which is then transformed into audio music. Ultimately, music conditioned on the input text and the transformed audio music is generated. The experiments demonstrate that the proposed method produces music more aligned with the input text compared to individual models. Moreover, the method proves particularly effective in generating music lasting between 30 to 74 seconds, and shows consistency improvement on individual models under variable input text lengths.