Spatial location from running binaural signals using cepstral analysis and logarithmic filtering

This paper proposes an acoustic model for predicting the acoustical room characteristics from a running binaural signal. This is accomplished via training a convolutional neural network on a precedence effect model to extract the spatial locations of the direct sound source and its early reflections. The precedence effect model extends and modifies the BICAM algorithm with cepstral analysis [Tyler, J., Si, M., \& Braasch, J. Acoust. Soc. Am. 151] and a logarithmic filter. The logarithmic filter takes human perception into account and provides better separation at higher frequencies. A synthetic dataset of binaural signals was generated using anechoic orchestral recordings with added reflections and reverberations. The binaural model generates binaural activity maps from binaural input signals, which are then used to train a convolutional neural network. The ability to predict the traits of a direct sound source and its reflections has applications in academic areas like perceptual modeling and room acoustical analysis. It can also be applied to industrial areas such as television and movies, video games, and augmented and virtual reality, to name a few. [Work supported by the National Science Foundation: HCC-1909229.]