You can't really do this with a spectral library because the pixels in the ROI will be converted to a mean spectrum so you will only have one sample for each class and there is no way to 'group' spectra into classes for a classification. Maximum likelihood and other classifiers that require the calculation of covariance will need to have multiple samples per class (number of bands + 1). A spectral library can only work with classifiers such as SAM or minimum distance. However, I think the best way to accomplish this is to use the API where you can calculate the statistics needed from the ROIs on one image and then apply those statistics to other images. They images must have the same number of bands to do this (so it matches what is in the covariance matrix). I would use ENVIROIStatisticsTask and then pass this to ENVIMaximumLikelihoodClassificationTask. If you prefer ENVI Classic API, this would be ENVI_STATS_DOIT and CLASS_DOIT.