A method is described that enables maximum-likelihood (ML) demodulation for MIMO communications over frequency-selective channels. An equalizer is typically employed to suppress inter-symbol interference (ISI) due to frequency-selectiveness of the channel, but the noise of the equalizer output can be highly correlated such that standard ML-MIMO demodulations cannot directly apply. The method comprises first constructing equivalent post-equalization MIMO channel and noise covariance matrix, and then de-correlating the equalizer output so that ML or near-ML MIMO demodulations can be applied to improve the performance. Additionally, successive ISI cancellation (SIC) is described for further performance improvement.