In this paper, we address the problem of channel estimation and signal detection in large MIMO FEC-coded systems assuming finite alphabet modulations. We consider a semi-blind iterative expectation maximization algorithm which relies on a limited number of pilot sequences to initialize the estimation process. We propose to include the estimation process within a turbo finite-alphabet simplicity (FAS)-based detection receiver. To that purpose we define two estimation updates from the FEC decoder output. Simulations carried out in both determined and undetermined configurations show that the resulting scheme outperforms the state-of-the-art receiver which uses an MMSE estimation criterion and that it reaches the maximum-likelihood lower-bound.