Can you compare HMM's with different number of hidden states?

Question

I'm using Hidden Markov Models (hmmlearn to name the package being used) to classify different fishing gear based on vessel trajectories. For that purpose I trained 3 different HMM's, one for each fishing gear using prelabeled data. I then use the score method of hmmlearn to get the log-likelihood for an unknown sequence on each of the 3 models. I then pick the model belonging to the highest likelihood. While 2 models consist of 2 hidden states, one probably has 3 hidden states.. Now to my question: Can I use this approach when utilizing HMM's having different number of hidden states?

I got good results for the first two models using a 70/30 split with all models having the same number of hidden states but the latter only resulted in a 50% accuracy so I want to change the parameters a bit.