Separarea oarbă a surselor sonore din mixturi binaurale / Blind source separation from binaural mixtures

The PhD thesis investigates a series of methods for separating voice sound sources from mixtures obtained by convolution with head related impulse responses or measured binaural impulse responses. In the first phase, methods for measuring and extracting head related impulse responses are developed, as well as methods for analyzing them by extracting the binaural cues. In order to take advantage of the spatial separation of the sources, a method for determining the direction of arrival of the sound sources based on the adaptive eigenvalue decomposition is researched. Next, four methods for separating sound sources from reverberant mixtures are investigated: the expectation maximization method, based on statistical grouping as a mixture of Gaussians of the spectrogram observations given by the mixing vectors, respectively the interaural level difference and the interaural phase difference; the variational Bayes method based on a new model for combining prior distributions for the mixing vectors and binaural cues; the multi-channel non-negative matrix factorization method based on the decomposition of the observations of the power spectrograms in the form of a matrix product which describes the time and frequency activities for the separated sources; independent low-rank matrix analysis based on the combination of independent vector analysis and non-negative matrix factorization. The methods are compared and analyzed from the performance point of view presenting the most suitable applications for each.