Описаны система превращения сигнала телерадиовещания в текст для украинского языка и моделирование особенностей, специфических для него – нерегулярность лексического ударения и высокая флективность. Разработанная система реализует подход клиент–сервер и позволяет просматривать пятиминутные сегменты речи синхронно с результатом распознавания речи.
Описано систему перетворення сигналу телерадіомовлення в текст для української мови та моделювання особливостей, специфічних для неї – нерегулярність лексичного наголосу та висока флективність. Розроблена система реалізує підхід клієнт–сервер і дає змогу переглядати п’ятихвилинні сегменти мовлення синхронно з результатом розпізнавання мови.
Introduction: Broadcast data processing is an important task for information society. The experience in development of real-time systems for Ukrainian dictation and speech record recognition on several computational platforms is the base for the described R&D devoted to extracting text from broadcast speech signal. Methods: The modeling is focused on features that are specific particularly for Ukrainian such as lexical stress and high inflexibility. 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 implemented a data-driven stress prediction algorithm that represents words as sequences of substrings and searches for one or more sequences with the best criteria. As a Slavonic language Ukrainian is highly inflective and tolerates relatively free word order, which motivates transition from word- to class-based statistical language model. Experimental research: Modeling both stressed and unstressed vowels leads to recognition accuracy improvement. Introduction word equivalence classes to the Language Model significantly decreases RAM consumption keeping the same recognition accuracy level. The developed experimental system implements client–server approach and allows for browsing 5-minute broadcast segments synchronously with speech recognition result. Conclusion: Language-specific speech feature modeling is beneficial for a speech recognition system. The created broadcast speech-to-text system opens news prospectives for broadcast stream analysis in Ukraine.
