For a single-speaker system, we divided the database of each speaker into 4 sets: 3 sets for the adaptation and one set for the test. We analyzed the performance of a recognition system for each speaker after successive adaptations ofthe adaptation database.