Reverberation Robust Feature Extraction for Sound Source Localization Using a Small-Sized Microphone Array

摘要

Conventional methods for sound source localization using microphone arrays are usually addressed from the signal processing viewpoint, where the sound source location is treated as a continuous parameter to be estimated over some spatial space. Actually, in some practical scenarios, such as in conference rooms and cars, sound source locations are only confined to some predefined areas. Therefore, it is more reasonable to deal with the problem from a machine learning point of view. By incorporating the prior information available about sound environments, machine learning-based methods have the potential to better deal with sound source localization in the presence of room reverberation. The key to machine learning-based sound source localization methods is how to extract effective source location features. The existing feature extraction schemes, such as the popular time-difference-of-arrival features, however, are not suitable for small-sized sensor arrays, due to the fact that sound source localization in reverberant environments become much challenging for small-sized arrays. To combat the problem, in this paper, we propose a reverberation robust feature extraction method for sound source localization based on sound intensity (SI) estimation using a small-sized microphone array. In particular, three robust feature extraction procedures have been employed in the proposed features, including normalization, phase transform weighting, and fully incorporating the redundancies in SI estimation. Simulation and real-world experimental results both show that the proposed sound source location features are more effective for small-sized arrays in reverberant environments when compared with the existing features.