Identification of multipole sources with neural deconvolution

Convention of the European Acoustics Association Forum Acusticum, pp. 2143–2150

Abstract

Beamforming techniques for localizing sound sources often assume the radiation of multiple monopoles and use a simplified transfer function between the source and the array.While this approximation provides satisfactory results in many practical cases, it reaches its limits for more complex sources.By considering multipole sources such as dipoles, we can improve the results; however, the orientation of the dipoles must be known a priori.In this paper, the neural deconvolution method, which employs neural networks to help deconvolve the beamforming map, is extended to include multipole sources.This is done by developing an extended version of the DAMAS algorithm, which we call DAMAS-MS, where MS stands for "Multiple-Sources", and by applying a neural grid compression technique to make the whole process run fast enough for real-time applications.We show that it is possible to obtain reasonable estimates of source strength, location and radiation type without prior knowledge of the dipole orientation, and fast enough for real-time applications.

Authors 3

1. Thiago Henrique Gomes Lobato corresponding

   HEAD Acoustics (Germany)

   Affiliation as printed

   HEAD acoustics GmbH

   HEAD acoustics GmbH, 52134 Herzogenrath, Germany

2. Roland Sottek

   HEAD Acoustics (Germany)

   Affiliation as printed

   HEAD acoustics GmbH

   HEAD acoustics GmbH, 52134 Herzogenrath, 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

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