Clustered speech and signal processing

This paper investigates the formation of ad-hoc microphone arrays for the purpose of recording multiple sound sources by clustering microphones spatially distributed within a room. A novel codebook-based unsupervised method for cluster formation using features derived from the Room Impulse Responses (RIRs) corresponding to each microphone is proposed and compared with baseline clustering and classification methods. The features correspond to the sequence of arrival time and time delays of echoes as estimated by peaks of the RIRs along with peak amplitudes. Results suggest that the proposed codebook based clustering algorithm can outperform KNN supervised classification method and kmeans unsupervised clustering method applied to microphone segmentation and clustering, in terms of clustering success rate and noise robustness.

A novel unsupervised multi-channel dereverberation approach in ad-hoc microphone arrays context based on removing microphones with relatively higher level of reverberation from the array and applying the dereverberation method on a subset of microphones with lower level of reverberation is proposed in this paper. This approach does not require any prior information about the number of microphones and their relative locations, however based on kurtosis of Linear Prediction (LP) residual signals, microphones located close to the active source are detected and utilized for the dereverberation process. The proposed method is a clustered enhancement method which can be applied with any dereverberation algorithm. The proposed method is not dependent on the recording setup so it requires no predefined threshold and it can be applied to unknown rooms with unseen speakers. Dereverberation results suggest that regardless of the applied dereverberation method, using a consciously chosen subset of microphones always yield better dereverberation results compared to blind use of all microphones.