Vector quantization applied to speech coding in the wireless environment
This thesis describes the development of the Voice Coding Development and Research (VoCoDeR) System, a software tool for testing and development of new speech coding methods. This tool enables a researcher to build a voice encoder and speech filters, determine an optimal bit allocation scheme, and/or create an error correction scheme, as desired. Using a channel simulation tool such as BERSIM, the user can create bit error patterns to corrupt the data and then decode the speech for playback and analysis. The system is based upon the North American Digital Cellular (NADC) 8kbps Vector-Sum Excited Linear Prediction (VSELP) speech coder and is currently capable of simulating the complete IS-54 source and channel coding scheme.
The system is tested using Multi-Stage Vector Quantization and Finite-State Vector Quantization (FSVQ) applied to the linear prediction coefficients. FSVQ provides significant bit rate savings over previous methods of quantization. A variety of coefficient representations are compared including log-area ratios, arcsine reflection coefficients, line spectrum pairs and immittance spectrum pairs. This has allowed the recently introduced immittance spectrum pairs to be tested using vector quantization. Multiple distortion measures are also examined. The VoCoDeR System provides a tool that will allow an engineer to work on new speech coding algorithms or to determine an optimal source and channel coding scheme.