Estimation of Late Reverberation Characteristics from a Single Two-Dimensional Environmental Image Using Convolutional Neural Networks

In augmented-reality (AR) applications, reproducing acoustic reverberation is essential to create an immersive audio experience. The audio component of an AR system should simulate the acoustics of the environment that is experienced by the users. Earlier, in virtual-reality (VR) applications, sound engineers could program all the reverberation parameters for a particular scene in advance or when the user is at a fixed position. However, adjusting the reverberation parameters using conventional procedures is difficult because the unlimited range of such parameters cannot be programmed for AR applications. Therefore, dynamically estimating the reverberation characteristics based on the environments in which the users move is necessary. Considering that skilled acoustic engineers can estimate the reverberation parameters using the images of a room without performing any measurements, we trained convolutional neural networks (CNNs) to estimate the reverberation parameters using two-dimensional images.