This paper explores the impact of the integration of technology and creativity in music on future music composition, with a focus on the design and application of music generation systems based on deep learning. First, a Markov chain model is constructed to analyse the distribution of musical notes. The CNN-Attention mechanism is then combined to extract the main melody, and an improved Transformer-XL model is used to enhance the quality of music generation. Objective evaluations show that in terms of repetition rate, the improved Transformer-XL model achieved significant optimisation, with a repetition rate of only 17.63%, representing a 48.82% decrease compared to Melody_LSTM. Subjective evaluations revealed that the system achieved an average score of 4.12 across five operational performance dimensions. In terms of music generation quality, the system scored 4.5 and 4.6 on the two key dimensions of style consistency and musical authenticity, respectively, demonstrating a clear advantage over the control system’s scores of 3.2 and 3.3.
