Sound Localization and Separation for Teleconferencing Systems

Nowadays, a table-top microphone system, which is used for teleconferencing, is usually installed in conferencing rooms. Using such conference phones, all active speakers are recorded simultaneously and a mixture of the speakers is transmitted. For convenience, it would be great to have a device which acquires remotely a high-quality speech signal for every single participant. So this thesis regards an appliance for remote acquisition of speech signals in common office environments. A combination of a microphone array and signal processing has been applied to localize and separate the speech contributions of the participants. Audio based localization is performed through a Steered Response Power Beamformer and smoothed through particle filtering. The separation process is based on Geometric Source Separation which joins the benefits of Beamforming and Blind Source Separation algorithms. Performance evaluations have shown that separation quality depends strongly on localization stability and accuracy. The audio-based localization detects the sound sources with a success rate of more than 80 percent the correct position with an accuracy of 4 degree, in an office environment. With these localization data a continuous separation of the speakers can be performed with a mean signal-to-interference ratio of more than 27 dB. All obtained results presented in this thesis show that reliable remote acquisition of speech signals is possible. In future, devices separating different speakers are entirely conceivable.     «

Nowadays, a table-top microphone system, which is used for teleconferencing, is usually installed in conferencing rooms. Using such conference phones, all active speakers are recorded simultaneously and a mixture of the speakers is transmitted. For convenience, it would be great to have a device which acquires remotely a high-quality speech signal for every single participant. So this thesis regards an appliance for remote acquisition of speech signals in common office environments. A combinatio...     »