Exploiting the Overtones in Online Localization of Sound Sources

Sammanfattning: Localization of multiple sound sources from a microphone array is a challenging task that has been a research topic for decades. The challenges stem from the diversity of acoustic contexts due to reverberation and disturbances. Therefore, recent approaches that impose structure on the sound sources, rather than on the noise, have shown promising results. The offline method HALO was proposed in 2016 for sparse localization of stationary sources by exploiting the overtone structure. In this thesis, an online recursive localization method is proposed inspired by a similar signal model. The proposed method consists of a two step procedure. First, the pitch estimator named PEACE estimates the fundamental pitches along with their harmonics in an adaptive dictionary. Secondly, the method named PLEASE estimates the positions of every estimated pitch. An adaptive scheme for setting the sparsity inducing regularization parameters in PEACE is also proposed by exploiting the spatial dynamics. As a result, the compound method PEACE-PLEASE is only left with physically meaningful user defined parameters that are trivial to set for a given application. The proposed method is compared to GCC-Phat on both simulated data and recordings from an anechoic chamber. The results indicate that PEACE-PLEASE outperforms GCC-Phat on both the anechoic dataset and the simulated data. At last, potential directions in research are highlighted and discussed.