Localisation and separation of simultaneous voices with two microphones

A new way of localising two simultaneous voices, called the f-p method, is introduced. Test results of investigations carried out in the free-field (i.e. on a flat roof with no obstructions) are presented. The f-p method applies a neural, N-tuple, adaptive windowing technique to new binary image representation of the phase information derived from two microphones. It is closely related to a psychoacoustic theory of the human interaural time difference localisation mechanism. The microphones need not be fixed for the purposes of speech enhancement and so the method has, in principle, the same spatial freedoms as the human system. A crucial factor in the success of the device is the presence of gaps in the spectrum of each source at different times and frequencies. The voices can, on the free-field case, be almost completely separated by using the localisation information. In the two-voice free-field case each voice is localised approximately seven times per second. The method involves no complex calculations and it is likely that it could be implemented in real-time, to allow tracking of moving sources, with a DSP board and a fast microprocessor. There is room for improvement in the values of the f-p parameters.<>