Degree Name

Master of Engineering (Hons.)


Department of Electrical and Computer Engineering


This thesis implements and develops the voice coding part of the U.S digital cellular mobile telephone system, the Vector-Sum Excited Linear Predictive Coder (VSELP). Given the specification from Electronic Industry Association, a simulation structure is fully established and documentation of the standard is thoroughly discussed, identifying some pitfalls in the standard. The performance of the simulated coder is shown to produce good quality synthetic speech having a typical SNR of around 10 dB, operating at 7950 bits per second. To ensure the correctness of the coder implementation, this thesis also design several methodologies for verifying the self excitation and the codebook excitation searching algorithms. The VSELP coder uses an efficient fixed point covariance lattice algorithm (FLAT) for filter coefficient estimation. The performance of the FLAT algorithm is compared with the Autocorrelation (AUTO) method using three measures. The comparison results show that the FLAT algorithm is superior to the AUTO for Linear Prediction analysis. This thesis explores the strategies of lowering its bit rate of the VSELP making it a candidate for low bit rate application such as mobile satellite communication. Various schemes have been designed to modify the standard VSELP. A modified VSELP coder which operates at 5167 bps giving a quality around 5.7 dB in terms of SNRSEG, is chosen as the best candidate in low bit rate coding comparable to the original version. Further, error sensitivity of each parameter of the modified VSELP is investigated using a simulated satellite channel, showing the LPC coefficient and the frame energy to be the most sensitive parameters. The burst error sensitivity of the modified VSELP coder has also been tested. To control channel errors when transmitting parameters, an interleave error masking mechanism operating without redundant bits is fully developed, migitating the effect of channel errors. The speech quality of the modified VSELP using the interleave error masking scheme shows around 2 dB SNRSEG improvement for the BER being 10^ compared to the speech without error protection