Maximum likelihood adaptation of histogram equalization with constraint for robust speech recognition