Kolloquium - Details zum Vortrag

Sie werden über Vorträge rechtzeitig per E-Mail informiert werden, wenn Sie den Newsletter des kommunikationstechnischen Kolloquiums abonnieren.

Master-Vortrag: Source Localization Using Binaural and Monaural Cues

Till Hardenbicker
Montag, 21. Oktober 2019
11:00 Uhr
Hörsaal 4G

Acoustic source localization exploiting binaural cues is used in applications such as binaural conferencing and hearing aids. Binaural cues are features that can easily be extracted from binaural signals but only provide an ambiguous location estimation, since they do not, for example, enable a front-back distinction. In the human auditory system, this distinction is made based on so-called monaural cues, which are signal features that arise from filtering effects of outer ear, head and torso.

This thesis deals with the full-sphere localization of speech sources based on binaural and monaural cues contained in binaural signals. It is demonstrated that unlike binaural cues, monaural cues provide information about the elevation angle. Attention is particularly paid to the extraction of monaural cues. As the original source signal is not known, this extraction is not trivial. Several algorithms are proposed to estimate the source position by identifying Head-Related Transfer Functions without knowledge of the source signal. The approaches can be divided into least-square and codebook-based concepts and artificial neural network-based concepts. Binaural and monaural cues are combined such that the monaural estimators only resolve the ambiguities left by the binaural estimators.

Different estimator topologies are individually optimized and then analyzed with regard to their performance and potential for standalone applications. The results indicate that all proposed estimators enable a front-back distinction. While one estimator performs best in exchange for a high computation time, two of them compromise well between computation time and localization accuracy.

zurück

Alle Interessierten sind herzlich eingeladen, eine Anmeldung ist nicht erforderlich.