| Much research has been published on source localization in a theoretical or laboratory setting. Source localization is very important for several applications, including beam-forming. However it has not been applied widely in practical situations. This work will examine the problems faced by source localization in real acoustical settings and will suggest an algorithm that can robustly estimate locations quickly. Additionally this work will propose a new algorithm that can localize multiple simultaneous sources.; Real audio environments contain significant reverberation and reflections that complicate localization of sources. Strong reflections can be more energetic than the signal arriving via the direct path. This adversely effects algorithms based on time delay estimates. However, more robust methods, like Steered Response Pattern - Phase Transform (SRP-PHAT), tend to be too expensive for the applications envisioned in this paper. This work will quantify the computational complexity of SRP-PHAT to search the space of possible locations.; An alternative algorithm, Hybrid Localization, will be presented that combines two different algorithms to achieve faster computation. It works very well for single source localization and can easily be extended to multiple sources. The multiple source extension makes use of pre-defined sub-arrays and a neural network. Hybrid Localization achieves this reduction in computation time without degrading localization robustness.; While researching Hybrid Localization, a new statistical model was developed to predict the resulting general cross-correlation (GCC) phase transform (PHAT). The model improves on existing models by changing assumptions on the source signals to better match reality. It can be used to quickly simulate the effectiveness of a room, a microphone array geometry, or a localization. In addition it can make predictions for a source in motion. |
