Channel estimation techniques for turbo multiple-input multiple-output receivers are addressed, which exploit a priori information to refine the poor performance of initial channel estimates solely using training symbols. We study linear minimum mean squared error channel estimation (LMMSE-CE) using second-order soft statistics to incorporate the degree of decoding reliability. Utilizing this additional information effectively minimizes the error propagation which the maximum likelihood based expectation maximization channel estimation (ML-CE) suffers. The LMMSE-CE provides a smaller mean square error of channel estimation and block error rate than the ML-CE.