Sparse Bit-Allocations Based on Partial Ordering Schemes With Application to Speech and Audio Coding

The majority of speech and audio encoders today operate at rates of less than 1-2 bit/input-sample. Reducing the rates even further is a challenge in particular for coders using waveform-based coding paradigms. Specifically for transform-based coders it is often the case that transform coefficients are quantized at less than 1 bit/coefficient (on average) in many areas of the spectrum. Dealing with such cases is particularly challenging when some coefficients are assumed to be independent identically distributed random variables with little underlying predictable structure. This prompts a study on how to improve encoding in such situations beyond simply increasing the dimension of quantizers, an option that may not be practical or perceptually relevant in some coder designs. This paper looks at a general class of schemes motivated by observations on statistical variations at low dimensions. These schemes, termed partial ordering schemes, show advantages objectively (in mean square error) and perceptually by targeting randomly positioned peaks in the spectrum. At low rates, such schemes create sparse bit allocations that allow for further optimization using random noise fill. A listening test is presented demonstrating the advantages of such an approach