A Style and Attention Mechanism Based Generative Adversarial Neural Network (Style-Atten GAN) for Flute Music Composition

Abstract

This study delves into the task of automated flute music composition by leveraging deep learning methods to automatically generate flute music with single and multiple tracks. To enhance musical complexity and diversity, we propose a new music generation model called Style-Atten GAN, with the integration of a generative adversarial network (GAN) model, self-attention mechanism, and stylistic characteristics of the flute. This directs the music generative process to focus on critical musical elements and structures in typical flute audios, thereby fostering the production of compositions that are not only novel, but also rich in musicality and expressiveness. Through a comparative experiment and analysis, this study evaluated the effectiveness of the proposed Style-Atten GAN model against recent music composition methods of the same type. Quantitative and qualitative analytical indicators highlight the superiority of the proposed neural network approach in producing stylistically diverse and creatively rich flute music. The findings reveal that GAN models, when augmented with the attention mechanism and unique stylized characteristics of the flute instrument, can outperform conventional methods in terms of fidelity, adherence to musical aesthetics, and the ability to capture the nuanced expressiveness of flute music.