A Machine Learning Approach to Detecting Sound-Source Elevation in Adverse Environments

摘要

Recent studies have shown that Deep Neural Networks (DNNs) are capable of detecting sound source azimuth direction in adverse environments to a high level of accuracy. This paper expands on these findings by presenting research which explores the use of DNNs in determining sound source elevation. A simple machine-learning system is presented which is capable of predicting source elevation to a relatively high degree of accuracy in both anechoic and reverberant environments. Speech signals spatialized across the front hemifield of the head arc used to train a feedforward neural network. The effectiveness of Gammatonc Filter Energies (GFEs) and the Cross-Correlation Function (CCF) in estimating elevation is investigated. Binaural cues such as Interaural Time Difference (ITD) and Interaural Level Difference (ILD) are also examined. Using a combination of these cues, it was found that source elevation to within 10° could be estimated to an accuracy of approximately 80% for both anechoic and reverberant environments.