Investigation of Split Quantization Properties for the Speech Spectrum Source

We investigate split quantization of the LSF source from an information theoretic perspective. It is a well known fact that split quantization is inferior to unconstrained quantization, due to the independent treatment of sub vectors. Here, we quantify lower bounds for the split loss, and suggest conditional quantization of splits as a method to reduce the losses. The investigation is based on information theoretic estimates of performance, rather than on the performance of actual quantization schemes. Simulations for a 10-dimensional LSF source point at a loss of around 8 bits if scalar quantization is performed, and a loss of approximately 2 bits for a 2-split quantization, both compared to 10 dimensional vector quantization. Furthermore, the results suggest that these losses can be considerably reduced using conditional quantization, down to as low as 0 to 2 bits. (Fredrik Nordén, Thomas Eriksson (Chalmers University of Technology))