Multimedia Systems Computer Networks and Communications

A multidimensional evaluation system for theme-based music generation using AE-VGAN modeling

Li, Peng, Zhang, Jingjing, Wu, Xiaojun, Jin, Fangzhu, Dong, Liren
Springer Nature · 2026-11-14 · DOI 10.1007/s00530-026-02770-4

Abstract

Recent advances in deep learning have substantially improved the quality of symbolic music generation, but reliable evaluation remains difficult. Existing metrics often focus on isolated statistical or music-theoretical properties and provide limited characterization of thematic preservation, structural development, and perceptual quality, which are particularly important in theme-based generation. This paper presents a multidimensional evaluation framework comprising 13 base metrics organized into three dimensions: music generation quality, thematic representativeness, and generation diversity, together with corresponding composite scores. The framework is examined through human listening tests and excerpt-level concurrent criterion-validity analysis, showing strong agreement between the proposed numerical quality metrics and perceptual judgments. To demonstrate its application, we further develop an Attention-Enhanced Variational Generative Adversarial Network (AE-VGAN), which integrates variational and adversarial learning with channel attention and self-attention mechanisms. Experiments on GiantMIDI-Piano, LAKH MIDI, and POP909 show that AE-VGAN achieves competitive performance across the proposed evaluation dimensions and the highest subjective overall-quality scores among the compared models on all three datasets. These results support the proposed framework as a systematic and perceptually grounded approach to evaluating theme-based symbolic MIDI generation under the studied Western classical and popular-music settings.

Read at publisher →