This paper constructs an efficient computational framework for style migration and melody generation. A decoding architecture from MIDI audio to CQT spectrogram is proposed based on the diffusion model, and a onedimensional U-Net structure is introduced to optimize the noise prediction process and improve the inference efficiency of the traditional diffusion model. Utilize VAE to map high-dimensional audio data to low-dimensional space to save computational cost. Design the conditional diffusion model based on cross-attention mechanism to realize high-fidelity migration of music styles. Propose a melody generation method based on LabVIEW random numbers to balance creative inspiration and melodic structural integrity. The model of this paper is applied to the practice of music style migration and melody generation to verify the practical value of the model. The results show that the quality of CQT spectrograms generated by the model in this paper is more than 80%, and the style migration rate is more than 90%. Using the piano roll window to visualize the music melody generated after style migration can enhance the melody intuition and the flexibility of segmentation adjustment. In the subjective evaluation, the generated melodies of this paper’s model get the best results in two dimensions: coherence and emotional expression, which can effectively realize the music style migration and high-quality melody generation.
