Speaker localization and separation in teleconferences

At the Institute for Data Processing at the Technical University of Munich, a lot of research effort is spent in the development of a conference phone. In this thesis the focus is on the recording, the speaker localization and the source separation. In the past, several algorithms were tested in the conference scenario, but there was no comparison of the different algorithms with the same test data. Firstly, a suitable microphone array for each of the three treated algorithms was designed and built with a 3D plotter. Afterwards, the impulse-responses for the three microphone arrays were estimated for 0° , 10° , 20° elevation and in 5° steps in azimuth to make a statement about the properties of the different microphone arrays. Furthermore, more than 1000 recordings were made in the audiolab and in the videolab to evaluate the different algorithms. As well as the subjective examination of the auditory impression the SIR, SAR and SDR-values were determined for all the recordings. From the results of the experiments, the best localization and the best separation algorithms were chosen, which will be used in the following processing steps, for example speaker recognition.     «

At the Institute for Data Processing at the Technical University of Munich, a lot of research effort is spent in the development of a conference phone. In this thesis the focus is on the recording, the speaker localization and the source separation. In the past, several algorithms were tested in the conference scenario, but there was no comparison of the different algorithms with the same test data. Firstly, a suitable microphone array for each of the three treated algorithms was designed and bu...     »