Fig. 1From: Robustness of ad hoc microphone clustering using speaker embeddings: evaluation under realistic and challenging scenariosScheme for clustering and cluster based source separation. Features (either Mod-MFCCs or speaker embeddings) extracted from the microphone signals are used to cluster the microphones. Inter- and intra-cluster information is then exploited to extract the sources dominant in each speech cluster. Yellow blocks indicate stages at which speaker separation can be performed — and which we use for evaluation. These consist of initial masking, delay and sum beamforming (DSB), fuzzy membership value aware DSB (FMVA-DSB) and postfiltering one of the DSB outputs. The dotted box is a condition that is not included in the tabulated resultsBack to article page