In the present work, a large database comprising a set of features related to geometry, type of boundary condition, type of material and experimental value of the first natural frequency for 184 historic masonry towers is collected. The database is then used to develop empirical formulas for the estimation of the fundamental frequency related to the first bending mode. The data are divided into a training set and a validation set for a correct assessment of the empirical equations, generated through power regression and artificial neural network. Different combinations of features are exploited as independent variables to investigate their influence on the dynamic behaviour of masonry towers. To improve the accuracy of the prediction, distinct models are generated for specific clusters with likely homogeneous behaviour, according to the boundary conditions (isolated or bounded) and type of materials (brick, stone or mixed masonry).