Deconvolution with neural grid compression: A method to accurately and quickly process beamforming results

The Journal of the Acoustical Society of America, vol. 153, pp. 2073

Abstract

Beamforming results depend on the spatial resolution of the microphone array used, which may lead to sources close to each other being considered as one. Deconvolution methods that consider all directions simultaneously, such as DAMAS, produce better results in these situations. However, they have a high computational cost, often lack sufficient speed to be used in real-time applications, and have limited accuracy at lower frequencies. This paper introduces a hybrid method to perform deconvolution using a neural network that can improve the speed of deconvolution on high-resolution grids by more than 2 orders of magnitude, while also generating sparser maps without sacrificing accuracy compared to the compressed DAMAS method.

Authors 3

1. Thiago Henrique Gomes Lobato

   HEAD Acoustics (Germany)

   Affiliation as printed

   HEAD acoustics GmbH 1 , Herzogenrath, 52134, Germany

2. Roland Sottek

   HEAD Acoustics (Germany)

   Affiliation as printed

   HEAD acoustics GmbH 1 , Herzogenrath, 52134, Germany

3. Michael Vorländer Aachen Institute for Hearing Technology and Acoustics

   RWTH Aachen University

   Affiliation as printed

   Institute for Hearing Technology and Acoustics, RWTH Aachen University 2 , Aachen, 52074, Germany