In this paper, we introduce an adaptation of a multivariate feature selection method to deal with functional features. In our case, observations are described by a set of functions defined over a common domain (e.g. a time interval). The feature selection method consists on combining variable weighting with a feature extraction projection. Although the employed method was primarily intended for observations described by vectors in ℝn, we propose a simple extension that allows us to select a set of functional features, which is well suited for classification. This study is complemented by the incorporation of Functional Principal Component Analysis (FPCA) that project functions into a finite dimensional space were we can perform classification easily. Another remarkable property of FPCA is that it can provide insight about the nature of the functional features. The proposed algorithms are tested on a pathological voice detection task. Two databases are considered: Massachusetts Eye and Ear Infirmary Voice Laboratory voice disorders database and Universidad Politécnica de Madrid voice database. As a result, we obtain a canonical function whose time average is enough to reach similar performances to the ones reported in the literature.