Texas Instruments controls entry to its corporate computer facility by speaker verification. During two and one half years of operation, 0.4 million verifications have been performed for 300 users. Current performance is 99.9% user acceptance with 99% impostor rejection. Initial user acceptance is not as good (99%), but steady‐state performance is achieved after 50–100 verifications. Careful user orientation is important, and user adaptation and voice consistency continue to improve over a long period of time. The variance of input speech data for those 36 users (24 men, 12 women) who had been verified at least three thousand times was plotted versns session number. Average variance was 108 at 250 sessions, 105 at 500, 101 at 1000, 96 at 2000, and 93 at 3000 sessions. How do you explain this continued improvement?