文献类型：期刊文献

英文题名：A comprehensive study on optimized neural networks for automatic piano music generation

作者：Zhang, Aimin[1];Tian, Runxuan[2]

第一作者：张爱民

通讯作者：Zhang, AM[1]

机构：[1]Beijing Union Univ, Coll Special Educ, Beijing, Peoples R China;[2]Univ New South Wales, Coll Comp Sci, Sydney, Australia

第一机构：北京联合大学特殊教育学院

通讯机构：[1]corresponding author), Beijing Union Univ, Coll Special Educ, Beijing, Peoples R China.|[1141759]北京联合大学特殊教育学院;[11417]北京联合大学;

年份：2025

卷号：19

期号：12

外文期刊名：SIGNAL IMAGE AND VIDEO PROCESSING

收录：;EI(收录号：20253619125141);Scopus(收录号：2-s2.0-105015067952);WOS:【SCI-EXPANDED(收录号:WOS:001567113700032)】；

基金：This study supported by 2021 Humanities and Social Sciences Project of the Ministry of Education of China, Vertical Project, Project No. 20210055.

语种：英文

外文关键词：Automatic piano music generation; Adjustable woodpecker optimization (AWO); Malleable generative adversarial network (MGAN); AI music composition; Multimedia signal processing

摘要：Piano music has long been a focus in automated music generation due to its complexity and versatility. However, optimizing neural networks for generating musically coherent, emotion-driven piano compositions remains a challenge. This study aims to develop an optimized neural network for automatic piano music generation using the adjustable woodpecker optimized malleable generative adversarial network (AW-MGAN). To achieve this, a dataset comprising a diverse collection of piano music pieces ranging from classical to modern styles was gathered. These pieces were selected based on their harmonic complexity, rhythmic diversity, and emotional expressiveness to ensure a rich source for training. In the context of multimedia signal processing, data preprocessing was conducted to reduce noise and simplify patterns within the music data while retaining essential features. Feature extraction was implemented using the Mel frequency cepstral coefficients (MFCC), a technique known for its effectiveness in capturing the essential characteristics of music. The processed data was then fed into the proposed AW-MGAN framework; the generative model was optimized through the AW optimization algorithm. This approach allows the model to adjust and refine its neural network layers dynamically, ensuring unnecessary complexity is minimized and essential features are retained. Proposed framework showed significant improvements in generating harmoniously coherent and emotionally resonant piano compositions. Research that we have contrasted with the approach and conventional technique evaluates average score (7.521), recall (95.2%), precision (90.65%), entropy (6.215). Results demonstrated measurable enhancements in both the diversity and precision of generated music, confirmed by both machine-based evaluation metrics and human listening tests.