Research on Polyphonic Music Generation Algorithm Based on GPT Large Model

Along with the rapid technological progress in the field of artificial intelligence, music generation algorithms based on large-scale pre-trained models have increasingly become the focus of academic attention. Existing polyphonic music generation techniques have limitations in terms of melodic complexity and harmonic diversity. To address the challenges of structural accuracy and long-range dependency modelling in polyphonic music generation, this paper proposes a targeted fine-tuning algorithm based on GPT macromodels, which improves the generation quality by incorporating domain-specific mechanisms. On the basis of GPT, the method embeds directional cross-track attention to enhance the modelling of vocal interactions, designs dynamic interval weight mask constraints and acoustic compliance, and introduces beat-phase embedding to enhance the temporal structure perception, forming a “GPT + domain-enhanced” generation framework. Experimental results show that the method demonstrates significant advantages in objective evaluation dimensions, such as note accuracy and harmonic consistency. The proposed method opens up an innovative path for polyphonic music composition and demonstrates the great potential of this technology in practical scenarios.