Hearing Places: Source Localization with Microphone Arrays

Abstract

Among the many applications of microphone array hardware is the ability to separate multiple simultaneous sources of sound. We apply this capability to a novel version of the well-studied source localization problem. This variation needs no assumptions about the acoustic properties of the room, nor that we know the number of simultaneous sources, instead relying on the singular assumption that speakers need only be localized if they reside at one of several known spots in the room. We explore three different machine learning-based approaches before settling on a final solution to the problem. The first approach is limited greatly in the types of signals it can localize, and generally performs with poor accuracy. The algorithms of the second approach can be used to localize many simultaneous sources, or detect when none are present, but no single algorithm succeeds at one of these cases without a severe drop-off in performance on the other. The final resulting algorithm localizes any number of sources with high accuracy (85%, on average). It is susceptible to more false positives as the number of simultaneous sources increases, but experiences no such trend in false negatives. It is robust to noise (15dB of SNR), though robustness also decreases with increasing numbers of simultaneous sources.