A system for large vocabulary continuous speech recognition of Slovenian language
is described. Two types of modelling units are examined: words and sub-words.
The data-driven algorithm is used to automatically obtain word decompositions. 
The performances of one-pass and two-pass decoding strategies were compared. 
The new models gave promising results.
The recognition accuracy was improved by 2.5% absolute at the same recognition time.
 On the other hand we
 achieved 30% increase in real time performance at the same recognition error.