Article: Music intelligent creation method based on LSTM and multi-scale attention Journal: International Journal of Arts and Technology (IJART) 2025 Vol.15 No.6 pp.1 - 25 Abstract: To address challenges in current AI music composition such as inconsistent styles, high melody repetition rates, and weak emotional expression, this study innovatively proposes a music intelligent creation method integrating long short-term memory (LSTM) with multi-scale attention mechanism (MAM). The model employs multi-layer LSTM to capture long-term dependencies in musical sequences, incorporates a residual module to optimise training processes and prevent gradient vanishing, and combines multi-scale attention mechanisms to dynamically weight features across different temporal scales including melody, rhythm, and harmony. These enhancements significantly improve the quality of generated music. Experimental results show that the rhythm consistency of the music creation model was 95.07% after 300 generations, the melodic beat matching reached 90.81%, the melodic repetition rate was only 6.23% after 100 generations, and the melodic accuracy reached 98.81%. The research model has high innovation and diversity in music creation tasks, and can generate rich and varied melodies, avoiding the dilemma of monotonous repetition. The research not only provides new technological means for music creation, but also promotes the application and development of AI in artistic creation, laying a solid foundation for future music creation. Inderscience Publishers - linking academia, business and industry through research

Title: Music intelligent creation method based on LSTM and multi-scale attention

Authors: Liping Li

Addresses: Conservatory of Music, Jilin Normal University, Siping, 136000, China

Abstract: To address challenges in current AI music composition such as inconsistent styles, high melody repetition rates, and weak emotional expression, this study innovatively proposes a music intelligent creation method integrating long short-term memory (LSTM) with multi-scale attention mechanism (MAM). The model employs multi-layer LSTM to capture long-term dependencies in musical sequences, incorporates a residual module to optimise training processes and prevent gradient vanishing, and combines multi-scale attention mechanisms to dynamically weight features across different temporal scales including melody, rhythm, and harmony. These enhancements significantly improve the quality of generated music. Experimental results show that the rhythm consistency of the music creation model was 95.07% after 300 generations, the melodic beat matching reached 90.81%, the melodic repetition rate was only 6.23% after 100 generations, and the melodic accuracy reached 98.81%. The research model has high innovation and diversity in music creation tasks, and can generate rich and varied melodies, avoiding the dilemma of monotonous repetition. The research not only provides new technological means for music creation, but also promotes the application and development of AI in artistic creation, laying a solid foundation for future music creation.

Keywords: intelligent creation; long short-term memory; LSTM; multi-scale attention; residual module; musical creation; multi-scale attention mechanism; MAM.

DOI: 10.1504/IJART.2025.150181

International Journal of Arts and Technology, 2025 Vol.15 No.6, pp.1 - 25

Received: 29 Aug 2025
Accepted: 17 Oct 2025
Published online: 02 Dec 2025 *

Title: Music intelligent creation method based on LSTM and multi-scale attention

Keep up-to-date