Sound Source Localization and Modeling: Spherical Harmonics Domain Approaches

Abstract

Sound source localization has been an important research topic in the acoustic signal processing community because of its wide use in many acoustic applications, including speech separation, speech enhancement, sound event detection, automatic speech recognition, automated camera steering, and virtual reality. In the recent decade, there is a growing interest in the research of sound source localization using higher-order microphone arrays, which are capable of recording and analyzing the soundfield over a target spatial area. This thesis studies a novel source feature called the relative harmonic coefficient, that easily estimated from the higher-order microphone measurements. This source feature has direct applications for sound source localization due to its sole dependence on the source position.

This thesis proposes two novel sound source localization algorithms using the relative harmonic coefficients: (i) a low-complexity single source localization approach that localizes the source' elevation and azimuth separately. This approach is also appliable to acoustic enhancement for the higher-order microphone array recordings; (ii) a semi-supervised multi-source localization algorithm in a noisy and reverberant environment. Although this approach uses a learning schema, it still has a strong potential to be implemented in practice because only a limited number of labeled measurements are required. However, this algorithm has an inherent limitation as it requires the availability of single-source components. Thus, it is unusable in scenarios where the original recordings have limited single-source components (e.g., multiple sources simultaneously active). To address this issue, we develop a novel MUSIC framework based approach that directly uses simultaneous multi-source recordings. This developed MUSIC approach uses robust measurements of relative sound pressure from the higher-order microphone and is shown to be more suitable in noisy environments than the traditional MUSIC method.

While the proposed approaches address the source localization problems, in practice, the broader problem of source localization has some more common challenges, which have received less attention. One such challenge is the common assumption of the sound sources being omnidirectional, which is hardly the case with a typical commercial loudspeaker. Therefore, in this thesis, we analyze the broader problem of analyzing directional characteristics of the commercial loudspeakers by deriving equivalent theoretical acoustic models. Several acoustic models are investigated, including plane waves decomposition, point source decomposition, and mixed source decomposition. We finally conduct extensive experimental examinations to see which acoustic model has more similar characteristics with commercial loudspeakers. Show more