Geometric-Wave Acoustic (GWA) dataset
收藏资源简介:
GWA数据集是由马里兰大学创建的大规模音频数据集,包含约200万合成房间脉冲响应(IRs)及其详细的结构和模拟配置。该数据集从超过6.8K高质量、多样化的专业设计房屋中采样声学环境,这些房屋以语义标记的3D网格表示。GWA数据集采用基于语义匹配的新型真实世界声学材料分配方案,使用句子转换器模型。数据集创建过程中,通过自动校准几何声学射线追踪与有限差分时域波解算器,计算高质量的脉冲响应,对应于准确的低频和高频波效应。GWA数据集的应用领域包括声传播和渲染、视频的深度音频合成以及使用深度学习的语音处理,旨在解决复杂场景中准确波声模拟的问题。
The GWA dataset is a large-scale audio dataset developed by the University of Maryland, encompassing approximately 2 million synthetic room impulse responses (IRs) along with their detailed structural and simulation configurations. Acoustic environments are sampled from over 6.8K high-quality, diverse professionally designed residences, which are represented as semantically tagged 3D meshes. The dataset adopts a novel real-world acoustic material assignment scheme based on semantic matching, leveraging Sentence Transformer models. During the dataset creation process, high-quality impulse responses are calculated via automatically calibrated geometric acoustic ray tracing and finite-difference time-domain (FDTD) wave solvers, which accurately capture low-frequency and high-frequency wave effects. The GWA dataset has applications in sound propagation and rendering, deep audio synthesis for videos, and deep learning-based speech processing, aiming to solve the challenge of accurate wave-based acoustic simulation in complex scenes.




