Doctor of Philosophy
Department of Electrical and Computer Engineering
Secker, Philip Lawrence, Channel-optimised trellis source coding for the memoryless Gaussian channel and applications to spectral parameter coding, Doctor of Philosophy thesis, Department of Electrical and Computer Engineering, University of Wollongong, 1994. http://ro.uow.edu.au/theses/1340
The trellis source coder is a high performance source coder that operates at relatively low complexity. Channel-optimised (CO) trellis source coding consists of a jointly designed source encoder and decoder for a given noisy channel. As such, the characteristics of the channel are an integral part of the overall design. This thesis examines various methods of operation on the additive white Gaussian noise (AWGN) channel and provides an application to speech spectral parameter coding.
To achieve good performance on the AWGN channel consideration must be given to using the continuous or real information provided by the channel. This work describes a number of systems that, variously, use different degrees of quantized channel information. If the decoder is constrained to accept information at the same rate that the encoder provides, it is apparent that the use of a-priori information can improve performance. While maximum a-posteriori (MAP) detection can considerably improve performance, such a system is not jointly designed. Hence a simple decision-feedback detector is proposed and a joint system is developed. Performance for the Gauss-Markov source is compared favourably against maximum likelihood (hard decision) and MAP detection.
The second system partitions the channel output space into four regions (symmetric about the origin) converting the channel into a binary input, 4-ary output discrete memoryless channel (DMC). The decoder operates directly with the soft-decision information. The third system is estimation based. The decoder is an optimum, non-linear estimator that accepts continuous information from the channel. This system marginally outperforms the previous system indicating that 4-level quantization realises most of the gain of the infinite-level estimator. Further improvements to this system are obtained by extending the scalar trellis to a two-dimensional vector trellis which also enables signalling in two dimensions. The extension to a vector alphabet (QPSK signalling) yields a quite reasonable improvement without increased encoder computational complexity. A system representing a joint design of trellis encoder, channel, estimator decoder and modulator is considered. By modifying the signal constellation the costs associated with a signal point may be related to its position in the plane. Both constant energy, variable phase average energy constraints were imposed with each further improving performance.
The Line Spectrum Pair (LSP) parameters of low bit-rate speech coders present an ideal vector source for which CO trellis source coding is highly suited. This primarily due to the strict bit-rate constraints, the high noise characteristics channel and the requirement of low complexity. The system is shown to offer good robustness to very noisy channels at 33 bits/frame. For operation on the AWGN channel it is shown that both MAP detection and estimator decoding provide a worthwhile gain.