用神经网络从房间参数预测声学响应，速度快且听感无差别。
Room Impulse Response Prediction with Neural Networks: From Energy Decay Curves to Perceptual Validation
- 输入房间尺寸、材料吸声率和位置，用神经网络预测能量衰减曲线
- 重建的混响响应与真实数据相关性超0.95，听觉测试无显著差异
- 适合需要快速生成逼真声场的沉浸式音频应用开发者
房间脉冲响应（RIR）预测对声学建模、空间音频和沉浸式应用至关重要，但传统仿真与测量成本高、耗时长。本文提出一种神经网络框架，仅凭房间尺寸、材料吸收系数及声源-接收器位置即可预测能量衰减曲线（EDC），并通过反向微分重构对应RIR。训练数据基于包含真实几何结构、频率相关吸声特性和多样化声源-接收器配置的声学仿真生成。客观评估采用均方根误差（RMSE）和自定义的EDC损失函数，以及相关性、均方误差（MSE）和频谱相似度指标衡量重建后的RIR性能。通过MUSHRA听觉测试进行感知验证，结果表明预测与参考RIR之间无显著感知差异。实验表明该框架可实现准确且感知可靠的RIR预测，为实际声学建模与音频渲染提供可扩展解决方案。
原文摘要 · Abstract (English)
Prediction of room impulse responses (RIRs) is essential for room acoustics, spatial audio, and immersive applications, yet conventional simulations and measurements remain computationally expensive and time-consuming. This work proposes a neural network framework that predicts energy decay curves (EDCs) from room dimensions, material absorption coefficients, and source-receiver positions, and reconstructs corresponding RIRs via reverse-differentiation. A large training dataset was generated using room acoustic simulations with realistic geometries, frequency-dependent absorption, and diverse source-receiver configurations. Objective evaluation employed root mean squared error (RMSE) and a custom loss for EDCs, as well as correlation, mean squared error (MSE), spectral similarity for reconstructed RIRs. Perceptual validation through a MUSHRA listening test confirmed no significant perceptual differences between predicted and reference RIRs. The results demonstrate that the proposed framework provides accurate and perceptually reliable RIR predictions, offering a scalable solution for practical acoustic modeling and audio rendering applications.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。