一种基于多特征融合与压缩激励模型的音乐主旋律提取算法  

A MUSICAL MELODY EXTRACTION ALGORITHM BASED ON MULTI-FEATURE FUSION AND SQUEEZE-EXCITATION MODEL

作　　者：何丽[1] 刘浩 He Li;Liu Hao(School of Information,North China University of Technology,Beijing 100144,China)

机构地区：[1]北方工业大学信息学院,北京100144

出　　处：《计算机应用与软件》2023年第5期160-166,261,共8页Computer Applications and Software

基　　金：国家自然科学基金项目(61972003,61672040)。

摘　　要：音乐旋律提取领域的研究普遍存在整体准确率(OA)难以提升、虚警率(VFA)高的问题。此外,在该领域应用深度学习方法时,还存在可用训练数据少、训练时间长的问题。针对以上问题,以语义分割模型(Segmentation)为基础,提出使用多特征融合的压缩-激励模型(SENet),以改进旋律提取的效果。将训练数据转换为GC和GCOS为原始数据,加入梅尔倒谱系数(MFCC)和色度特征(Chroma Feature)。为进一步发挥多特征融合的优点,将SENet中的压缩-激励模块(SEBlock)加入Segmentation模型中。实验表明,加入特征可以提升音高准确率、音级准确率(RPA,RCA),并且收敛速度提升,使用70%的数据便可接近基线算法的效果;加入SEBlock后可以在进一步提升准确率的同时降低虚警率,更好地发挥多特征融合的优势。Researches in the field of music melody extraction generally have the problems that the overall accuracy rate(OA)is difficult to improve and the false alarm rate(VFA)is high.In addition,when applying deep learning methods in this field,there are also problems of less available training data and longer training time.To solve the above problems,based on the semantic segmentation model,this paper proposes a multi-feature fusion squeeze-excitation model(SENet)to improve the effect of melody extraction.We converted the training data into GC and GCOS as the original data,and added the Mel Cepstral coefficient(MFCC)and chroma feature.In order to further exploit the advantages of multi-feature fusion,we added the compression-excitation module(SEBlock)in SEnet to the Segmentation model.Experiments show that adding features can improve pitch accuracy and chroma accuracy(RPA,RCA),and the convergence speed is improved.It achieves the effect of close to the baseline algorithm by using 70%training data of origin data.After adding SEBlock,it can further improve accuracy while reducing the false alarm rate,and the advantage of multi-feature fusion can be better utilized.

关 键 词：语义分割 多特征融合 压缩-激励模型 梅尔倒谱系数 色度特征 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]