The use of biometrics for authentication mechanisms is becoming more and more important for research as well as industry. Keystroke dynamics is a biometric authentication method that improves the security of password-based applications. The performance of biometric keystroke recognition is still an open research issue. In fact, the extracted features relevant to a personal way of typing become less representative over time. This can lead to a failure in the biometric verification task. Because of the changes of such features, the representative model has always to be updated. In this paper, we use growing and sliding windows as template update methods based on a statistical classifier. We also demonstrate that user-specific thresholds, varying from an update session to another, allows to reduce the error rates compared to the update with a fixed threshold.