Confronting HMM-based phone labelling with human evaluation of speech production

The paper presents outcomes of an experimental study in which various modes of HMM labelling were tested on two groups of Czech speakers differing in the quality of their performance. Apart from the highest precision for the given speaking style - in this case read connected speech - we were also looking for indications that the HMM labeller might function differently for the group of good speakers and for the group of poor speakers. It turned out that impressionistically evaluated good and poor speakers were, at least in some of the modes, labelled with a different degree of precision.