基于融合特征与卷积神经网络的耳鸣偏好音自动标注研究  

Research on automatic tagging of tinnitus preference music based on fusion features and convolutional neural network

作　　者：徐瑞阳 何培宇[1] 方安成 冯楚楠 潘帆[1] XU Ruiyang;HE Peiyu;FANG Ancheng;FENG Chunan;PAN Fan(College of Electronics and Information Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电子信息学院,成都610065

出　　处：《生物医学工程研究》2024年第4期338-343,共6页Journal Of Biomedical Engineering Research

基　　金：四川省自然科学基金资助项目(2022NSFSC0799)。

摘　　要：在耳鸣声治疗中加入患者偏好音乐可有效改善耳鸣治疗效果。为满足患者个性化偏好的音乐需求,本研究提出了一种基于logMel与Hpcp融合特征的音乐自动标注新方法。该方法首先通过提取音乐的声学与乐理相关特征,然后输入双分支EfficientNetV2-s网络,进行音乐标注。经测试,本研究方法在MTAT数据集上的ROC-AUC值达到了0.9119,相较于其他音乐标注方法,标注性能有一定提升,对耳鸣偏好音选择具有一定的参考价值。Adding the patient's preferred music to tinnitus treatment can effectively improve the effect of tinnitus treatment.To meet the personalized music preferences of patients,we proposed a novel method for music automatic tagging based on the fusion of logMel and Hpcp features.Firstly the acoustic and music theory related features of the music were extracted and input to the two-branch EfficientNetV2-s network for music annotation.The tested results showed that the ROC-AUC value of this method on the MTAT dataset reached 0.9119.Compared with other music annotation methods,the annotation performance has been improved to some extent.This study has a certain reference value for tinnitus preference tone selection.

关 键 词：耳鸣治疗 音乐标注 双分支结构 特征融合 注意力机制 深度学习 

分 类 号：R318[医药卫生—生物医学工程] R764.45[医药卫生—基础医学] TP183[自动化与计算机技术—控制理论与控制工程]