Speaker localization in conferencing systems employing phase features and wavelet transform

Some of existing conference system employ a distant microphone array instead of microphones dedicated for each user. This approach is much more convenient although suffers from much higher noise sensitivity. One of the possible solutions is employing beamforming techniques to focus on the user that is speaking at the moment. However, beamformer needs information about the direction of arrival (DOA) parameter which is usually provided by analysing the phase differences between signals. Effectiveness of such solution decrease dramatically when the environment becomes noisy. In this paper, a novel, robust meetings diarization system is described. The decision about which user is speaking at the moment is based not only on spacial features of signal (i.e., speaker´s localization) but also on spectral features. The microphone array estimates speaker localization employing generalized cross-correlation with phase transform (GCC-PHAT). Additionally, the speaker recognition system which employs wavelet-Fourier transform (WFT) extracts spectral features of voice. Described solution is much more robust than the one basing on speaker recognition or speaker localization only. The experiments during meetings in regular meeting room show that it is less noise sensitive and the switching between speakers is several times faster.