This paper describes a real-time speech recognition system for
Ukrainian designed basically for text dictation purpose
targeting moderate computation requirements. The research is
focused on features which are specific particularly for
Ukrainian. Given arguments confirm the necessity to
distinguish stressed and unstressed vowels in the phoneme
alphabet. Lexical stress irregularity implies expert
involvement for stress assignment. To automate this procedure
we propose a data-driven stress prediction algorithm that
represents words as sequences of substrings. The formulated
criteria that validates a substring sequence is based on a set of
words with manually pointed stresses and a large text corpus.
The described search algorithm finds one or more sequences
with the best criteria. As a Slavonic language Ukrainian is
highly inflective and tolerates relatively free word order.
These features motivates transition from word- to class-based
statistical language model. According to our experimental
research, 4-gram class-based LM occupies less space and has
promising prospectives. We describe a speech-to-texå web-
service where the proposed techniques are used as well as
several tools developed to visualize HMMs, to predict word
stress, and to manage cluster-based language modeling.