This paper presents a novel channel estimator for GSM space-time coded (STC) systems. It is based on applying the maximum likelihood (ML) principle not only over a known training sequence, but also over the unknown symbols. Performing the so-called turbo equalization at reception we can obtain very accurate estimations of the a posteriori probabilities of these unknown symbols. These probabilities play a key role in the computation of the ML channel estimator. We also show how the expectation-maximization (EM) algorithm can be used to solve the resulting channel estimation optimization problem. As a practical application we study how much can the training sequence length be reduced in a GSM system operating in subway environments.