Automatic speech recognition methods have been reviewed in the paper. Hidden Markov models for acoustic modeling were used, a language model and dictionary were created. The models were trained with recorded news, total 5859 words, for 2 hours and 20 minutes. Two phonetic systems (with 85 and 32 phonemes respectively) were developed on the basis of a Lithuanian pronunciation dictionary. The created models were integrated in a stenography app prototype for transcribing voice recordings. The testing results of the acoustic models showed a 94% accuracy for phrase recognition.