Tracking speakers in an audio stream

Audio information is processed to identify potential segment boundaries, corresponding to a speaker changes 220. Thereafter, homogeneous segments (generally corresponding to the same speaker) are clustered 230, and a cluster identifier is assigned to each identified segment. A segmentation subroutine identifies potential segment boundaries using the BIC model selection criterion. A window selection scheme considers a relatively small amount of data in areas where new boundaries are very likely to occur, and the window size is increased when boundaries are not very likely to occur. When a segment boundary is found in a window, the next window begins after the detected boundary, using the minimal window size. BIC tests can be eliminated when they correspond to locations where the detection of a boundary is very unlikely.