Modeling of Pronunciation, Language and Nonverbal Units at Conversational Russian Speech Recognition

The main problems of a conversational Russian speech recognition system development are variability of pronunciation, free word-order in sentences and presence of speech disfluencies. In the paper, pronunciation variability is modeled by creation of multiple word transcriptions. A syntacticstatistical language model that takes into account long-distant word dependencies is proposed for Russian language modeling. Also in this paper the results of analysis of such speech disfluencies as artefacts and filled pauses, which were extracted during segmentation of the Russian speech corpus, are presented. The recognition accuracy of nonverbal elements in the collected corpus was 87%. The proposed methods of pronunciation variability modeling and syntactic-statistical language model creation were realized in the software complex for Russian speech recognition. The performed experiments with large vocabulary using syntactic-statistical language model showed that word error rate of the system was 33%.