基于全连接神经网络的音乐厅音质分级评价  

Acoustic quality evaluation of concert halls based on fully connected neural network

作　　者：闫琛 李运江[1] 许华华 范波[1] YAN Chen;LI Yunjiang;XU Huahua;FAN Bo(Institute of Building Technology,Hubei Engineering Technology Research Center for Disaster Prevention and Mitigation,China Three Gorges University,Yichang 443002,Hubei,China)

机构地区：[1]三峡大学建筑技术科学研究所,湖北省防灾减灾工程技术研究中心,湖北宜昌443002

出　　处：《声学技术》2025年第2期276-283,共8页Technical Acoustics

摘　　要：为了准确和快速地利用所获得的声学客观参量对音乐厅音质进行评价,文章采用了全连接神经网络(fully connected neural network,FCNN)方法来构建音乐厅音质分级评价模型,探索了音乐厅设计和音质优化的新方法。文章将音乐厅的3类13种声学客观参量与音乐厅音质效果等级作为输入和输出,用于训练FCNN模型,得到了较精准的音乐厅音质分级评价模型。经过训练的音乐厅音质分级评价模型能够以决定系数R2=1的精度来对音乐厅音质进行分级评价。相较于传统的音质分级评价方法,基于FCNN的音质分级评价方法计算耗时约为前者的1/10。在此基础上,通过分析FCNN模型中输入层到隐藏层的权重矩阵,同时结合基于机器学习的决策树算法,文章对13种声学客观参量进行了权重优选,最终确定了影响音乐厅音质效果等级的声学客观参量排序。排序结果表明,在音乐厅音质评价中,时间类声学客观参量的权重明显高于其他类声学客观参量的权重。研究结果表明在实际音乐厅音质评价过程中使用FCNN方法可以减少传统分级评价方法导致的主观性误差影响,该方法可为优化音乐厅设计和提升听众体验提供支持。A fully connected neural network(FCNN)approach is utilized in this study to construct a grading evaluation model for concert hall acoustics,which can evaluate their quality accurately and efficiently with acquired acoustic objective parameters.By inputting three categories and thirteen types of acoustic objective parameters of concert halls,along with the corresponding quality grades outputs,a FCNN model was trained resulting in a highly accurate grading evaluation model for concert hall acoustics.The trained concert hall acoustic quality grading evaluation model is demonstrated exceptional precision in evaluating the grading with a coefficient of determination(R2)of 1.Compared to traditional grading evaluation methods,the computation time of the FCNN-based grading evaluation method is about one-tenth of the former.Additionally,with analysis of weight matrices between the input layer and the hidden layer,and decision tree algorithms based on machine learning,the weighted optimization of 13 acoustic objective parameters is conducted.Ultimately,significant acoustic parameters that impact the acoustic quality grades of concert halls were identified and ranked.In the ranked results of acoustic quality evaluation of concert halls,category of temporal acoustic parameters exhibit noticeably higher weightage compared to other categories.The research results show that FCNN method can minimize the influence of subjectivity of traditional grading methods.It can be applied in optimizing concert hall design and enhancing the audience experience.

关 键 词：音乐厅音质分级评价 机器学习 全连接神经网络 音乐厅参数化设计 音乐厅声学客观参量 

分 类 号：TU112[建筑科学—建筑理论] TB5[理学—物理]