In this paper, we accomplish matching score level fusion of multi-biometrics. In order to solve the incomparability among different classifiers’ outputs, Adaptive Confidence Transform (ACT) is introduced to convert the raw outputs of different classifiers to the estimates of posteriori probabilities conforming to different users. These posteriori probabilities are then combined using several fusion methods. Experiments conducted on a database (including face, iris, online signature and offline signature traits) of about 100 users indicate that for the same fusion method, ACT based normalization generally results in better verification performance and is more robust compared to other normalization methods. Effects of different normalization and fusion methods on combination of "strong" and "weak" classifiers are also examined.