Acoustic Speaker Localization with Strong Reverberation and Adaptive Feature Filtering with a Bayes RFS Framework

Abstract

The thesis investigates the challenges of speaker localization in presence of strong reverberation, multi-speaker tracking, and multi-feature multi-speaker state filtering, using sound recordings from microphones. Novel reverberation-robust speaker localization algorithms are derived from the signal and room acoustics models. A multi-speaker tracking filter and a multi-feature multi-speaker state filter are developed based upon the generalized labeled multi-Bernoulli random finite set framework. Experiments and comparative studies have verified and demonstrated the benefits of the proposed methods.