I'm using the LibriSpeech s5 recipe to train an acoustic model on reverberant speech using sentences in the train-clean-100 dataset (which I convolved with room impulse responses). Papers I've read (e.g. the ASpIRE paper) suggest training a GMM-HMM based system to perform forced alignment on the anechoic sentences, and then training a neural network based acoustic model on the corresponding reverberant sentences using the clean alignments as labels. I'd like to do the same but use a GMM-HMM acoustic model instead of a neural network.
Is it possible to train a GMM-HMM acoustic model on reverberant sentences using alignments generated from a GMM-HMM forced aligner trained on the corresponding anechoic sentences? If so, how would I accomplish this?