Speech parameter data reduction for recognition purposes