Informative features having important discriminatory power, also called high-level features, for classification of bearings faults can be automatically extracted from the spectrum of the vibration signal envelope without the a prior knowledge of the characteristic bearing frequencies. It was shown in a previous work that the Principal Component Analysis (PCA), when applied on a specific spectral matrix based on these spectral features., allows discriminating accurately between different faults conditions of bearings. Healthy and faulty bearings with faults on the outer-race, the inner-race and the balls are separated into distinct classes irrespective of the system operating point. The classification does not need any complex classifier like neural networks or support vector machines. There are still some difficulties, however, to discriminate between different levels of severity related to the faults in the bearing balls. The present work uses Linear Discriminant Analysis (LDA) to improve the classification of faults on balls according to their severity level, while only relying on the information carried out by the already used spectral features. Experimental results show that the LDA, besides its simplicity, extracts from the spectral features new variables having more discriminatory power than the principal components. The accuracy of the discrimination into the PCA and LDA spaces is evaluated using Bhattacharyya distance, a well-known measure of class separability. The linear discriminant axes allow for a good discrimination between different sizes of faults in bearing balls. The obtained results validate the contribution of the LDA space to the diagnosis of faults in bearings elements based on the proposed spectral features.