基于音乐情感的ERoPE-Transformer音乐生成方法  

ERoPE-Transformer:An approach to Music Generation Based on Musical Emotions

作　　者：张玉梅[1,3] 陈章杰 吕小姣 延成岭 卢恒 ZHANG Yu-mei;CHEN Zhang-jie;LV Xiao-jiao;YAN Cheng-ling;LU Heng(School of Computer Sciences;School of Journalism and Communication,Shaanxi Normal University,Xi’an 710062,China;Key Laboratory of Folk Song Intelligent Computing and Service Technology,Ministry of Culture and Tourism,Xi’an 710062,China)

机构地区：[1]陕西师范大学计算机科学学院,陕西西安710062 [2]陕西师范大学新闻与传播学院,陕西西安710062 [3]民歌智能计算与服务技术文化旅游部重点实验室,陕西西安710062

出　　处：《榆林学院学报》2025年第2期78-86,共9页Journal of Yulin University

基　　金：陕西省创新能力支撑计划项目(2022TD-26);国家社会科学青年基金项目(24CTQ016)。

摘　　要：针对音乐生成缺乏情感表达和交互性的问题,提出一种基于音乐情感的ERoPE-Transformer音乐生成方法。基于情感的旋转位置编码Transformer模型(Emotion Rotary Position Embedding Transformer,ERoPE-Transformer)以CP Transformer模型为基础模型,在CP编码的基础上加入情感标签,将音乐序列转换成离散的符号序列,并且引入旋转位置编码,提高模型的外推性,更好地捕获长序列建模中的音乐信息和情感特征。最后通过客观和主观实验对音乐质量和情感效果进行验证,与其他常用的音乐生成模型进行比较,包括Transformer-XL模型、CP Transformer模型和CEG-Transformer模型。客观实验从音高、节奏和音乐结构三个方面评价音乐质量,同时采用DUPSO-DSKSVM民歌快速分类算法对音乐情感进行分类,以验证生成音乐的情感效果。主观实验通过人耳听力测试从不同方面对音乐质量进行打分评价,同时使用人耳听力测试对音乐的情感效果进行评价。实验结果表明,本文提出的ERoPE-Transformer模型在音乐质量和情感表达上都有很好的效果,优于其他三种方法。To address the lack of emotional expression and interactivity in music generation,this paper proposes an ERoPE-Transformer music generation me thod based on musical emotion.Building upon the CP Transformer model,the Emotion Rotary Position Embedding Transformer(ERoPE-Transformer)incorporates emotional labels into CP encoding,converting music sequences into discrete symbolic sequences.By introducing rotary position encoding,the model enhances extrapolation capabilities and better captures both musical information and emotional features in long-sequence modeling.Experimental validation through objective and subjective assessments compares the proposed model with other common music generation approaches,including Transformer-XL,CP Transformer,and CEG-Transformer.Objective evaluations assess music quality through pitch,rhythm,and structural analysis,while the DUPSO-DSKSVM folk song rapid classification algorithm verifies emotional effects.Subjective experiments employ human auditory tests to evaluate both musical quality and emotional expression across multiple dimensions.Results demonstrate that the ERoPE-Transformer achieves superior performance in both music quality and emotional representation compared to the three baseline methods.

关 键 词：情感音乐生成 Transformer网络 旋转位置编码 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]