间接优化的高效Monte Carlo声传播研究

Realtime sound propagation has important application in the rapidly developing virtual reality and AI voice recognition domains. Geometry acoustic algorithms based on Monte Carlo (MC) path tracing are the most important realtime sound propagation methods, and they differ from the visual rendering counterpart in that the path sample distribution in the sampling domain cannot be directly controlled and the optimization of the sample distribution relies on indirect sampling mechanisms. This project aims at improving the quality/performance efficiency of MC sound propagation. To solve the key scientific problems of efficient MC sound propagation, including the qualitative metrics and their perceptual relevance, time domain indirect optimization of path sample distribution, error-controlled space-time reuse of path samples, the project proposes to start from studying the metric built on the impulse energy response that can be directly used to drive optimization algorithms. Based on the metric, frame coherence and the correlation between specific algorithm parameters and path length can be exploited to develop indirect sample distribution optimization algorithms. Space-time hierarchy of sound sources can be established, based on which time varying sample reuse algorithms can be proposed. The project also theoretically analyzes the convergence of the proposed algorithms. The project will significantly enhance the efficiency of sound propagation, and lay the theoretical and algorithmic foundation for the application of realtime sound propagation in consumer level VR and voice recognition.

实时声传播模拟在迅猛发展的虚拟现实和人工智能语音识别领域具有重要应用，而基于几何声学的Monte Carlo路径跟踪是其中最重要的一类方法。声传播的路径跟踪和面向图形绘制的路径跟踪的一个重要区别在于无法直接控制样本在目标域的分布，因此其质量/性能效率的改善依赖于间接样本优化分布方法。本课题面向高效Monte Carlo声传播模拟，围绕定量质量评价标准及其感知关联度、路径空间采样时域间接优化分布和误差控制的路径样本时空域重用等科学问题，研究基于冲激能量响应的信号特征建立可直接驱动声传播算法优化的质量评价标准，并在此基础上研究利用帧间连续性和特定算法参数与路径长度之间的内在关联性实现时域的间接样本分布优化，以及基于误差上限估计的声源时空域层次结构和在此之上的样本重用，并分析算法收敛性。课题有望大幅提高声传播模拟效率，为其在消费级虚拟现实和语音识别的高效样本合成中的应用和普及奠定理论和算法基础。

针对现有的声传播质量标准ERSNR不能直接反映人类对声音质量的感知的问题，我们提出了基于感知的声传播路径跟踪评价标准，这一新标准冲激响应和最终用户听到的声音的感知质量关联起来，从而使得可以在此基础上，根据最终对声音质量的要求自适应的的确定采样数量，从而在不对人耳感知的声音质量产生明显影响的情况下提升计算性能。采样不足是制约声音模拟质量的最重要的因素之一，我们又进一步的提出了引入深度神经网络来对低采样的声传播模拟结果进行质量提升的方法，并分别分析了针对冲激响应函数、冲激响应频谱和输出音频的质量优化方法。结合这两方面的技术，我们实现了预定的技术指标，在一个89个声源的复杂动态声学场景（54,000面片）中实现了25帧每秒的实时声传播模拟，性能相比现有算法提高5倍以上。这为实时声传播模拟在移动平台的虚拟现实等应用中的使用提供了新的思路。