Models based on the hyper-geometric distribution for estimating the number of residual software faults are proposed. The application of the basic model shows that its fit to real data is good. Two ways of improving the model, using a segmentation technique and composite estimation, respectively, are shown. The segmentation technique appears quite effective, particularly when the growth curve of the cumulative number of detected faults bends sharply. The applications of these models to real data are demonstrated.>