Demixing of Speech Mixtures and Enhancement of Noisy Speech Using ADRess Algorithm

This paper describes the ability of the Azimuth Discrimination and Resynthesis algorithm (ADRess) to separate multiple speech signals from two mixtures in a simulation environment. ADRess exploits the spatial signature of each of the contributing speech sources to demix the mixtures. Speech sentences taken from the TIMIT database and noise signals from the NOISEX database were mixed synthetically to create pairs of mixtures. ADRess can exploit the spatial signature of noise and speech sources to remove or isolate them from a mixture. To simulate the spatial location of different sources the relative attenuation and phase difference of each source between the two mixtures were manipulated. This was performed for numerous different angles of arrival so as to robustly test the algorithm. Objective measures and promising informal listening test results show the suitability of ADRess for cleaning noisy speech mixtures and document the performance of ADRess for speech mixtures with different numbers of sources