Semenov V. Development of adaptive speech enhancement methods based on autoregressive model of vocal tract

In the thesis a novel methods of enhancement of acoustic speech signals based on the autoregressive (AR) model of speech generation are developed. Concepts of physiological acoustics, methods of estimation theory, theory of stochastic processes, decision analysis were used. A noise-robust effective approach to estimation of vocal tract AR parameters based on application of spectral templates (quantums) is developed. Effective method of blind deconvolution of speech at noise background based on line spectral frequencies (LSF) is proposed. Method of LSF calculation which has advantages over existing approaches is developed. Efficient algorithm of block Kalman filtering is developed. It provides both improvement of segmental signal-to-noise ratio and reduction of computations in comparison with traditional methods. Developed in thesis methods were experimentally tested on real speech signals and at present time are applied in speech coding devices. Other areas of implementation are systems of telephony, speech recognition, storage of voice information, acoustical diagnostics.