Fuzzy-clustering-supported Assignment of Smart-Speaker-based Microphone Arrays to Acoustic Sources in Reverberant Acoustic Environments

Inhalt:
In realistic acoustic environments, the signal of interest may be severely disturbed, especially when a variety of sound sources are simultaneously active and the signal is picked up by a microphone far from the source. It is therefore of interest to assign microphone arrays of an ad-hoc acoustic sensor network to specific sources such that these can be acquired with high fidelity. The method proposed in this paper employs filter-and-sum beamformers with fixed equiangular look directions on several uniform circular arrays to generate a diverse set of signals. DNN-based highlevel representations of the beamformers’ output signals can then be utilized to assign specific beams to sources and to generate signal clusters for further processing. We evaluate the utility of the proposed method via an SIRbased measure and by automatic speech recognition on the beamformer’s output signal which is selected from the set of beams assigned to a specific source. Compared to the baseline case which uses a single beamformer signal, we observe that the proposed method leads to a notable improvement in the SIR and a reduction of word error rates.