The purpose of the thesis is the performance improvement of existing Automatic Speech Recognition (ASR) Engines for languages with a complicated grammar, like Russian or German, by the use of hybrid approaches for Language and Lexical Modelling.

There are a lot of different speech recognition engines, including open source software, to enable creating individual ASR-systems. Such systems operate quite well only if we train the acoustic model and compose the dictionary and language model properly. However, for languages with a complicated grammar the quality of such ASR-system is much worse. This especially applies when there are a lot of variations of one word, according to the different situations. In some languages (like in Russian) these words may differ by only one single letter. In this thesis some approaches will be applied for Language and Lexical Modelling in order to improve the ASR-system results for such languages.