On Robust LPC-spectrum Coding and Vector Quantization
Abstract: Digital speech coding is concerned with representing a discrete-time speech signal with as few bits per second as possible without introducing annoying artifacts in the coded speech. Many modern speech coders are based on Linear Predictive Coding (LPC) which employ a source-filter model for speech production. The filter is obtained by performing linear prediction analysis on the input signal and is updated on a frame basis. The process of quantizing the LPC-filter to a finite number of bits per frame is known as LPC-spectrum coding or LPC-spectrum quantization. The LPC-filter models the short-term spectral envelope of the speech signal which is a crucial factor for the perceptual quality of the coded speech. Hence, the LPC-spectrum coding must be carefully performed.Early LPC-spectrum quantization schemes employed scalar quantization whereas vector quantization is common today. Vector quantization (VQ) is an efficient lossy data compression technique that is a vital part of many modern source coders. VQ, however, suffers from problems due to complexity and robustness. The complexity grows exponentially with the number of bits. The sensitivity against errors when transmitting the VQ indices across a noisy channel increases with the efficiency of the VQ. Various constrained VQ schemes have been studied to overcome the complexity problems whereas issues such as index assignment and channel matched quantization have been studied to improve the channel robustness. This thesis is mainly devoted to the LPC-spectrum coding problem. Both scalar and vector quantization are studied. Various aspects of LPC-spectrum coding are assessed including tree-search of scalar and segmented vector quantizers for differential Line Spectrum Pairs, utilizing cepstral coefficients in multi-stage VQ, and voicing-specific LPC-spectrum quantization. The Linear Mapping by a Block Code VQ (LMBC-VQ) is a major contribution of this thesis. Block code vectors from a binary block code in systematic form are linearly mapped to the source space to obtain the reconstruction vectors of the VQ. The LMBC approach offers a number of advantages by providing a relation between index bits and the reconstruction vectors through mapping properties. In particular we control the channel noise influence. Channel robustness theory is presented along with design methods that gives VQs with inherent good robustness against channel noise. Simulation results for LPC-spectrum coding and for Gauss-Markov sources are shown. Several other desirable properties of the LMBC-VQ are also demonstrated.
This dissertation MIGHT be available in PDF-format. Check this page to see if it is available for download.