In this paper, a new joint symbol detection and channel estimation algorithm is proposed for MIMO-OFDM systems over frequency-selective fading channels using the variational Bayesian expectation-maximization (VBEM) algorithm. Since the VBEM algorithm can provide distribution-estimates of the parameters, the statistical information about the channel uncertainty is exploited to improve the evaluation of the extrinsic information in the soft-input soft-output detector which identifies the significant symbol combinations via list sphere decoder. In addition, two channel estimators are derived based on the posterior distributions of the transmitted symbols which are obtained from the space-time detection. The VBEM iterations are embedded in the turbo-processing of the receiver. Simulation results demonstrate that the proposed VBEM algorithm has more robust performance over the conventional EM techniques.