Advancements in Transformer-Based Music Generation: Exploring Applications in Personalized Composition and Music Therapy

Abstract

The use of Transformer models in music generation has greatly contributed to the development of automatic music generation, to produce MIDI sequences with long-range dependencies. These models, which are particularly effective for processing sequential information, outperformed such models as RNNs in terms of better representation of long-term dependencies across compositions. Prominent sources, like Dewangan, Singh, and Verma’s work on developing Musical sequences, etc., stress the importance of new training methodologies that are free from distortions to structures, but also to creativity. Furthermore, more recent adaptations of such methods are the Music Transformer (Huang etal., 2019) and its circuitous extension, the Multitrack Music Transformer (Dong etal., 2022), which have embedded more complex structures of music into Transformers for better coverage of human-like concepts. Ha and colleagues suggest a technique they refer to as Compositional Steering in a work published in 2022 and another technique called Stylistic Clustering introduced by Zhang and colleagues in 2024. Thus, there are still issues to be addressed: the tendency of AI-generated music to be emotionally shallow, and the problems of using AI in music creation. This research considers these changes, seeking to examine the effectiveness of Transformer models in improving the efficacy of music therapy and psychological treatments with the use of individualized music creation. Investigates the interplay between AI and human creativity in composing and utilizing music for both artistic and therapeutic purposes.