Binaural Sound Source Localisation in Complex Conditions

Abstract

There has been a growing interest in the reproduction of human spatial hearing behaviours, arising from the development of spatial audio signal processing techniques. To accurately localise single or multiple sound sources using humanoid apparatus, it is essential to be able to exploit the spatial-related features of the human subject filtering effect, which requires an understanding of both the feature characteristics and the mapping relationship to the source locations. In this thesis, we analyse and evaluate the localisation feature characteristics of binaural signal, and explore a method for constructing a localisation mapping model.

As a result of the reflecting and diffracting of human-like apparatus, sound waves are filtered before being captured by the eardrum, and the filtering effects result in various behaviours in the frequency domain. This thesis first summarises the characteristics of those behaviours and evaluates their importance to localisation. We analyse and evaluate the correlation between source location and three main interaural cues, which are interaural level differences, interaural time difference and interaural phase difference. Then, we explore the process to exploit those features using, and develop a novel feature vector by combining the most valuable spectra. Following this, by employing mutual information as the evaluation metric for frequencies selection, we propose a new feature location mapping model that embeds the feature evaluation process. The new mapping uses a multiple-tree structured model based on the random forest that shows high tolerance to noise. Through computational simulations and practical experiments, the model presents an improvement in both accuracy and robustness according to the comparison of the angular error and localisation correct rate. Finally, by combining our localisation method with the recent proposed direct path transfer function estimation method based on a convolutive transfer function model, we design a binaural localisation system for an unknown environment.

The remainder of this thesis demonstrates the possibility of using the active localisation cues in a binaural system. Based on observations of human active head rotation behaviour, we investigate the effect of dynamic features in binaural localisation. The analysis shows that head rotation enriches the variation of localisation features, which resolve the problem of cone-of-confusion and simplifies vertical-wise localisation in a 3-D space. In addition, we develop a multiple-source localisation method based on the head rotation process, which indicates that dynamic features would be the solution to many localisation problems caused by the limitation on the number of signal channels.