The lecture presented current convergence of communication systems and networks and the concept of automated services. Enabling technologies encompasse above all speech technology systems, such as: speech recognition and synthesis systems, spoken dialogue systems, and speech-to-speech translation systems. Speech technology systems were discussed in more details and based on case-studies their implementation in real life environment was discussed.
B.04 Guest lecture
COBISS.SI-ID: 14030102The authors focus on speech recognition and machine translation. Inflective languages require different approaches from that applied to English. Inflectional languages create words by adding inflectional morphemes to the base form. Consequently large set of words arises leading to data sparseness problem. In the article the authors discuss language dependent and language independent methods proposed by research community to solve or abate this problem.
B.06 Other
COBISS.SI-ID: 12498710The lectures were introduction to two topics: language modelling and machine translation. We discussed probabilities in language and give answer to the question why using natural language is easy for a human and difficult for a computer. Human resolve many ambiguities in language, but they are very hard problem for computers. We give mathematical theory of statistical language modeling. Language models are not self-sufficient. They are used as knowledge source of an user-oriented application. We show how language model can be used as a knowledge source of machine translation system.
B.04 Guest lecture
COBISS.SI-ID: 12990230Member of editorial board International journal of speech technology that publishes articles in the field of speech technology development and applications.
C.06 Editorial board membership
COBISS.SI-ID: 16846341The authors discussed the influence of hangover and hangbefore criteria on speech recognition accuracy. To define how many frames should be taken for hangover and hangbefore criteria and to decide if hangover and hangbefore criteria will be used, the duration of vowels, diphthongs and semi vowels was used. Frames with low spectral energy were detected as silence. Speech recognition experiments showed that hangover and hangbefore criteria can improve speech recognition accuracy.
B.03 Paper at an international scientific conference
COBISS.SI-ID: 13315350