Acoustic Source Localisation in constrained environments

Acoustic Source Localisation (ASL) is a problem with real-world applications across multiple domains, from smart assistants to acoustic detection and tracking. And yet, despite the level of attention in recent years, a technique for rapid and robust ASL remains elusive - not least in the constrained environments in which such techniques are most likely to be deployed. In this work, we seek to address some of these current limitations by presenting improvements to the ASL method for three commonly encountered constraints: the number and configuration of sensors; the limited signal sampling potentially available; and the nature and volume of training data required to accurately estimate Direction of Arrival (DOA) when deploying a particular supervised machine learning technique. In regard to the number and configuration of sensors, we find that accuracy can be maintained at state-of-the-art levels, Steered Response Power (SRP), while reducing computation sixfold, based on direct optimisation of well known ASL formulations. Moreover, we find that the circular microphone configuration is the least desirable as it yields the highest localisation error. In regard to signal sampling, we demonstrate that the computer vision inspired algorithm presented in this work, which extracts selected keypoints from the signal spectrogram, and uses them to select signal samples, outperforms an audio fingerprinting baseline while maintaining a compression ratio of 40:1. In regard to the training data employed in machine learning ASL techniques, we show that the use of music training data yields an improvement of 19% against a noise data baseline while maintaining accuracy using only 25% of the training data, while training with speech as opposed to noise improves DOA estimation by an average of 17%, outperforming the Generalised Cross-Correlation technique by 125% in scenarios in which the test and training acoustic environments are matched.