Shariy T. The Information Technology of Speech Data Processing Based on the Fuzzy Cognitive Models

The thesis is devoted to the actual scientific task of accuracy improvement for the automatic word recognition in speech signals. The current problems in the design of automatic speech recognition systems are analyzed. The alternative speech processing scheme that considers multilevel acoustic information is proposed. A new method of automatic speech signal segmentation that takes into account the cepstral smoothness measure along with the spectral transition measure is developed and experimentally grounded. The approach of weighting the speech segments on basis of signal prosodic parameters and taking the weights into account at the post-processing stage is proposed. The sound image model of a phoneme on basis of Karhunen-Loeve transformation is proposed. The FCAS fuzzy cognitive model of speech signal post-processing is developed. The model includes the network of feature-level, phoneme-level and word-level elementary phonetic processors. The correct word is recognized according to the output values of the processors. It is established that the proposed models and methods allow reducing the word error rate. The information technology and the CogniSPEECH bundled software for speech command recognition and keywords search in files are developed. The system characteristics are examined. It is showed that the system can be used in voice dialing and robot voice-control programs.