Although the subject of fusion is well studied, the effects of normalisation prior to fusion are somewhat less well investigated. In this study, four normalisation techniques and six commonly used fusion classifiers were examined. Based on 24 (fusion classifiers) as a result of pairing the normalisation techniques and classifiers applied on 32 fusion data sets, 4x6x32 means 768 fusion experiments were carried out on the XM2VTS score-level fusion benchmark database, it can be concluded that trainable fusion classifiers are potentially useful. It is found that some classifiers are very sensitive (in terms of Half Total Error Rate) to normalisation techniques such as Weighted sum with weights optimised using Fisher-ratio and Decision Template. The mean fusion operator and user-specific linear weight combination are relative less sensitive. It is also found that Support Vector Machines and Gaussian Mixture Model are the least sensitive to different normalisation techniques, while achieving the best generalisation performance. For these two techniques, score normalisation is unnecessary prior to fusion.