Automatic creation of real-time, emotion-based piano music pieces remains a challenge for deep learning models. While Generative Adversarial Networks (GANs) have shown promise, existing methods can struggle with generating musically coherent pieces and often require complex manual configuration. This paper proposes a novel model called Learning Automata-based Self-Attention Generative Adversarial Network (LA-SAGAN) to address these limitations. The proposed model uses a Generative Adversarial Network (GAN), combined with Self-Attention (SA) mechanism to reach this goal. The benefits of using SA modules in GAN architecture is twofold: First, SA mechanism results in generating music pieces with homogenous structure, which means long-distance dependencies in generated outputs are considered. Second, the SA mechanism utilizes the emotional features of the input to produce output pieces. This results in generating music pieces with desired genre or theme. In order to control the complexity of the proposed model, and optimize its structure, a set of Learning Automata (LA) models have been used to determine the activity state of each SA module. To do this, an iterative algorithm based on cooperation of LAs is introduced which optimizes the model by deactivating unnecessary SA modules. The efficiency of the proposed model in generating piano music has been evaluated. Evaluations demonstrate LA-SAGAN’s effectiveness: at least 14.47% improvement in entropy (diversity) and improvements in precision (at least 2.47%) and recall (at least 2.13%). Moreover, human evaluation confirms superior musical coherence and adherence to emotional cues.
Full Text Finder