基于神经网络的声学参数预测方法研究
Research of acoustic parameter prediction method based on neural network
万宇鹏 1周远波 2文捷 1陈政 1赵晶1
作者信息
- 1. 中国测试技术研究院声学研究所,四川成都 610021
- 2. 四川海岩声学科技有限公司,四川成都 610599
- 折叠
摘要
为更加准确高效地预测建筑声学客观音质参数,该文基于机器学习的室内中频混响时间和语言传输指数的神经网络预测方法.将基于机器学习的神经网络技术与计算机声学模拟仿真技术相结合,提取 800个厅堂建筑的10个典型特征参数和 3个目标参数,利用Odeon声学仿真平台,针对不同音质参数指标建立多个数值矩阵训练样本数据库,采用机器学习理论对混响时间、语言传输指数等指标进行BP神经网络数据拟合训练.对训练结果的均方误差、误差分布及回归系数进行评估,结果显示混响时间参数的训练均方误差小于 0.05 s,语言传输指数参数的训练均方误差小于 1.5×10-4,所有目标参数的回归系数R值均优于 0.95.评估结果表明,该神经网络具备良好的预测准确性、数据泛化性和应用适用性.经实例验证,依托该神经网络编译和封装的应用程序可以实现对目标参数的快速评价,减少人力物力,提高工作效率.
Abstract
In order to predict the objective sound quality parameters of architectural acoustics more accurately and efficiently,this paper studies the neural network prediction method of indoor mid-frequency reverberation time and speech transmission index based on machine learning.In this study,the neural network technology based on machine learning and computer acoustic simulation technology were combined to extract 10 typical characteristic parameters and 3 target parameters of 800 hall buildings.By using Odeon acoustic simulation platform,multiple numerical matrix training sample databases were established for different parameters of sound quality.Using machine learning theory,BP neural network data fitting training is conducted for reverberation time,speech transmission index and other indicators.The mean square error,error distribution and regression coefficient of the training results are evaluated.The results show that the training mean square error of reverberation time parameter is less than 0.05 s,and the training mean square error of speech transmission index parameter is less than 1.5×10-4,the regression coefficient R values of all target parameters are better than 0.95.The evaluation results indicate that,the neural network has good prediction accuracy,data generalization and application applicability.The application program compiled and packaged based on the neural network can realize the rapid evaluation of target parameters,reduce manpower and material resources,and improve work efficiency.
关键词
机器学习/神经网络/声学设计/参数预测/混响时间/语言清晰度Key words
machine learning/neural network/acoustic design/parameter prediction/reverberation time/speech intelligibility引用本文复制引用
基金项目
2023年度重要技术标准研究项目(ZYBZ2023-5)
出版年
2024
NETL
NSTL
万方数据