Supervised Fusion Music Composition Through Long Short-Term Memory and Stochastic Modelling

Abstract

Music composition has witnessed significant advancements with the infusion of artificial intelligence, particularly using Long Short-Term Memory (LSTM) networks. However, most existing algorithms offer minimal control to composers in influencing the genre fusion process, thereby potentially undermining their creative preferences. This study introduces a novel, two-phase algorithm for personalized fusion music generation that reflects the composer's individual preferences. In the first phase, melodies are generated for individual genres using Recurrent Neural Networks (RNNs) employing techniques like Sequential, Dense, and one-hot encoding. These generated melodies serve as input for the second phase, where an LSTM network fuses them into a coherent composition. Notably, the algorithm incorporates weights set by the composer for each genre, allowing for a personalized composition. A stochastic approach is employed in both phases to introduce creative variance while balancing structural coherence. We demonstrate this balance through various metrics offering a more tailored fused music generation experience enriched by stochastic modeling.