Article
Deep-Learning-based Sound Source Localization (en)
* Presenting author
Abstract:
Deep learning methods have proven and established themselves in many areas of spatial and immersive audio in recent years. One notable application is sound source localization, where deep learning models can compete with or even surpass classical approaches, especially in challenging acoustic conditions.However, it has been shown that deep learning approaches primarily achieve good and reliable results when the models are not used as black boxes that process vast amounts of raw data and autonomously learn implicit features application, but when application, model, and input data are finely tuned to each other. This pre-colloquium aims to provide a comprehensive yet accessible overview of some core components of deep-learning-based sound source localization using examples from recent research of the Institute of Communications Technology at the Leibniz University Hannover, including data and input feature preprocessing or model and problem design. In addition, the current research project "Hooray" will be introduced, in which the possibilities and boundaries of deep-learning-based sound source localization using a head-mounted microphone array are being investigated.