NLTK HMMTrainer unsupervised learning

Question

I want to check how to do unsupervised learning for HMM with the nltk HMM-Trainer. As far as i know, a HMM with supervised and unsupervised training should perform better than only supervised trained. I got following Code

#!/bin/python3
import nltk
from nltk.corpus import brown
from nltk.util import unique_list

def hmmTrainer(sents):
    tag_set = unique_list(tag for sent in sents for (word,tag) in sent)
    symbols = unique_list(word for sent in sents for (word,tag) in sent)
    return nltk.tag.HiddenMarkovModelTrainer(tag_set, symbols)

if __name__ == '__main__':
    corpus=brown
    print(len(corpus.sents()))
> 57340
    train= corpus.tagged_sents()[:10000]
    devel=corpus.tagged_sents()[10000:11000]
    test=corpus.tagged_sents()[11000:13000]

    trainer=tag.hmmTrainer(train+devel)
    hmm=trainer.train_supervised(train)
    print(hmm.evaluate(test))
> 0.3531
    tmp=[nltk.untag(sent) for sent in devel]
    hmm=trainer.train(labeled_sequences=train,
                  unlabeled_sequences=[tmp])
    print(hmm.evaluate(test))
> 0.0866

# this needs ~40h for training
    train= corpus.tagged_sents()[:30000]
    devel=corpus.tagged_sents()[30000:33000]
    test=corpus.tagged_sents()[33000:36000]

    trainer=tag.hmmTrainer(train+devel)
    hmm=trainer.train_supervised(train)
    print(hmm.evaluate(test))
> 0.59785
    tmp=[nltk.untag(sent) for sent in devel]
    hmm=trainer.train(labeled_sequences=train,
                  unlabeled_sequences=[tmp])
    print(hmm.evaluate(test))
> 0.1057

Whats my mistake?