Data-driven Threshold Selection for Direct Path Dominance Test


Creative Commons License

Olgun O., Hacıhabiboğlu H.

23rd International Congress on Acoustics, Aachen, Germany, 9 - 13 September 2019, pp.3313-3320

  • Publication Type: Conference Paper / Full Text
  • Doi Number: 10.18154/rwth-conv-238971
  • City: Aachen
  • Country: Germany
  • Page Numbers: pp.3313-3320

Abstract

Direction-of-arrival estimation methods, when used with recordings made in enclosures are negatively affected by the reflections and reverberation in that enclosure. Direct path dominance (DPD) test was proposed as a pre-processing stage which can provide better DOA estimates by selecting only the time-frequency bins with a single dominant sound source component prior to DOA estimation, thereby reducing the total computational cost. DPD test involves selecting bins for which the ratio of the two largest singular values of the local spatial correlation matrix is above a threshold. The selection of this threshold is typically carried out in an ad hoc manner, which hinders the generalisation of this approach. This selection method also potentially increases the total computational cost or reduces the accuracy of DOA estimation. We propose a DPD test threshold selection method based on a data-driven statistical model. The model is based on the approximation of the singular value ratio distribution of the spatial correlation matrices as a generalised Pareto distribution and allows selecting time-frequency bins based on their probability of occurrence. We demonstrate the application of this threshold selection method via emulations using acoustic impulse responses measured in a highly reverberant room with a rigid spherical microphone array.