On Using Meta-Learning to Overcome Challenges in Speaker Localization

IEEE EUROCON - International Conference on Smart Technologies (EUROCON), vol. 70, pp. 291–294

Abstract

Speaker localization machine learning models can suffer from poor performance if the training dataset samples are very few and not sufficient. Moreover, these models require that the deployment environments be very similar to those used to train them; otherwise, they will experience a noticeable degradation in their performance. In this paper, we address this issue and try to solve it using meta-learning approach. In addition, we introduce a performance comparison between machine-learning and non-machine-learning-based methods used to localize speakers in different environments and scenarios. Our method was put to the test in a real-life scenario for a typical intensive care unit (ICU) room.

Authors 3

1. Mahdi Barhoush Aachen INDA Research Area

   RWTH Aachen University

   Affiliation as printed

   RWTH University,INDA Research Area,Aachen,Germany

   INDA Research Area, RWTH University, Aachen, Germany

2. Ahmed Hallawah Aachen Department of Intensive Care and Intermediate Care

   RWTH Aachen University · Universitätsklinikum Aachen

   Affiliation as printed

   University Hospital RWTH Aachen,Department of Intensive Care and Intermediate Care,Aachen,Germany

   Department of Intensive Care and Intermediate Care, University Hospital RWTH Aachen, Aachen, Germany

3. Anke Schmeink Aachen INDA Research Area

   RWTH Aachen University

   Affiliation as printed

   RWTH University,INDA Research Area,Aachen,Germany

   INDA Research Area, RWTH University, Aachen, Germany