An algorithm for separating overlapping voices

The objective is to separate overlapping sounds i.e. to achieve the `cocktail party effect´. The main intended application is to realise an advanced hearing aid, but another possible use is as a front-end preprocessor for a speech recognition system. An algorithm has been developed and tests have been undertaken on seven subjects with normal hearing, in which the task is to recognise key words in a short sentence that is masked by a much stronger sentence from another speaker. In all cases an improved word score was obtained when the ratio of signal to interference was -12 dB. In many cases the improvement was very striking