Deep Learning Approach to Simultaneously Localize Acoustic Source and Receiver with a Single Room Impulse Response

This report proposes a method of simultaneously estimating the locations of an acoustic source and receiver in a three-dimensional shoe box-shaped room. We use a convolutional neural network-based model where the only input is a single room impulse response with reverberations. We also propose a method of handling the case of degeneracy using a small amount of information about the positions of the source and receiver relative to one another. In contrast to existing methods, we require no additional information or constraints. The model was shown to have effectively learned patterns in the room impulse response signal, achieving average error of 1.401 m which is shown to be better than a random guess. Although error was still large, we view this work as a proof-of-concept, and we expect that future modifications to the model architecture will improve accuracy substantially, and preprocessing the synthetic data used as the model's input would allow it to generalize better to real-world data sets.