The paper describes the information technology of speech recognition on TV news 
through thematic linguistic models. When building recognition system 
initially conducted separate texts for linguistic model into topics 
and build some specific dictionaries for each theme. In recognition 
using two-pass approach, where the first step keywords are searched 
for all subjects and determined message subject. Then on the second 
pass speech recognition is using a particular thematic linguistic model. 
Experiments on speech recognition with TV news channel NewsOne are conducted