Frequency-Independent Ambisonics Upscaling Using Deep Learning

IEEE International Conference on Acoustics Speech and Signal Processing, pp. 4586–4590

Abstract

Due to its flexibility, the higher-order Ambisonics (HOA) sound format is widely used in spatial audio applications like virtual reality (VR). The spatial accuracy is sensitive to the HOA order, which is limited by the recording and reproduction hardware. To enhance accuracy, the directional audio coding (DirAC) method prioritizes the improvement of perceptual quality over physically accurate spatial representations. We propose a novel neural network architecture that performs HOA upscaling within frequency subbands independently by leveraging the properties of spherical harmonics basis functions that are inherent in the HOA transformations. Although the proposed network is trained on synthetic data, a 63% variance reduction of the spatial similarity is achieved compared to DirAC, while both methods exhibit similar average performance for real-world measurement data. This makes our model beneficial for applications which require reliable spatial representations.

Authors 2

1. Egke Chatzimoustafa Aachen Institute of Communication Systems

   RWTH Aachen University

   Affiliation as printed

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

2. Peter Jax Aachen Institute of Communication Systems

   RWTH Aachen University

   Affiliation as printed

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