We propose a generic on-line (also sometimes called adaptive or recursive) version of the expectation-maximization (EM) algorithm applicable to latent variable models of independent observations. Compared with the algorithm of Titterington, this approach is more directly connected to the usual...