T-WGAN: A Transformer-Wasserstein GAN Approach for Melody Generation with Structural and Rhythmic Fidelity

Abstract

To address the current challenges of deep learning music generation models in capturing long-range dependencies, ensuring generation diversity, and maintaining training stability, this study proposes an optimized music melody generation model—T-WGAN. The model deeply integrates Transformer and Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP). This study first preprocesses the large-scale Lakh MIDI dataset, extracts single-track main melodies, and converts them into symbolic sequences using a REMI-based event representation. On this basis, the model innovatively adopts a generator based on Transformer decoder to learn the long-range structure of melodies. It also uses a critic based on Transformer encoder for stable adversarial training under the WGAN-GP framework to enhance the diversity and authenticity of generated melodies. Experimental results show that T-WGAN performs excellently on multiple key evaluation metrics. T-WGAN achieves a Rhythmic Consistency Rate (RCR) of 85.17%, significantly higher than baseline models (e.g., Transformer’s 75.68%). Its score on Fréchet Distance for Music (FDM) drops to 31.02, proving that the generated melodies are closer to real music in feature distribution. The conclusion indicates that the proposed T-WGAN model successfully addresses the three core issues in melody generation—structural integrity, diversity, and training stability—synergistically. The findings provide an effective technical approach for generating high-quality music melodies with both structural logic and innovation.