Higher-Order Ambisonics Upscaling Using Gated Recurrent Units

European Signal Processing Conference (EUSIPCO), pp. 171–175

Abstract

Higher-order Ambisonics (HOA) offer a flexible way to represent 3D sound field information, which makes them suitable for many applications, e.g., virtual reality (VR) and teleconferencing. However, the HOA order which dictates spatial accuracy is practically constrained by the number of microphones and loudspeakers. This work aims to increase the accuracy of the sound field representation by predicting missing HOA coefficients for higher orders. To achieve this, an existing deep learning-based upscaling method utilizes fully connected feedforward neural networks. Our novel approach replaces these fully-connected structures with gated recurrent units (GRUs), which allow to better leverage spatio-temporal dependencies inherent in HOA coefficients. Simulation experiments show that when trained under similar conditions, the proposed model outperforms the previous one by achieving lower mean squared error (MSE) between target and predicted HOA coefficients across various upscaling orders. In further experiments, we train the proposed model on synthetic sinusoidal data and evaluate the performance on test sets of complex real-world recordings. The superior performance of the proposed model in these experiments indicates its value in scenarios where obtaining real acoustic scene data with high orders is impractical.

Authors 2

1. Egke Chatzimoustafa Aachen Institute of Communication Systems

   RWTH Aachen University

   Affiliation as printed

   Institute of Communication Systems (IKS), RWTH Aachen University,Germany

2. Peter Jax Aachen Institute of Communication Systems

   RWTH Aachen University

   Affiliation as printed

   Institute of Communication Systems (IKS), RWTH Aachen University,Germany