A simulated "cocktail party" with up to three sound sources

Listeners identified spoken words, letters, and numbers and the spatial location of these utterances in three listening conditions as a function of the number of simultaneously presented utterances. The three listening conditions were a normal listening condition, in which the sounds were presented over seven possible loudspeakers to a listener seated in a sound-deadened listening room; a one-headphone listening condition, in which a single microphone that was placed in the listening room delivered the sounds to a single headphone worn by the listener in a remote room; and a stationary KEMAR listening condition, in which binaural recordings from an acoustic manikin placed in the listening room were delivered to a listener in the remote room. The listeners were presented one, two, or three simultaneous utterances. The results show that utterance identification was better in the normal listening condition than in the one-headphone condition, with the KEMAR listening condition yielding intermediate levels of performance. However, the differences between listening in the normal and in the one-headphone conditions were much smaller when two, rather than three, utterances were presented at a time. Localization performance was good for both the normal and the KEMAR listening conditions and at chance for the one-headphone condition. The results suggest that binaural processing is probably more important for solving the "cocktail party" problem when there are more than two concurrent sound sources.