Ultra Low Bit-Rate Coders

In this chapter, we present the definition and principles of ultra-low bit-rate coders. Here the emphasis is to point to the fact that this class of coders is typically the ‘vocoders’, which are ‘parametric’ coders that are essentially linear-prediction (LP) based vocoders. This is in contrast to the ‘waveform’ coders, which operate at the higher bit-rates. Among the various frameworks employed for realizing ultra low bit-rate speech coding, we restrict our attention to the generic linear-prediction (LP) based vocoding, (of which the LPC-10 coder [T82] sets the baseline) mainly for its rather ubiquitous adoption to a range of segment quantization vocoders as well as for its simplicity and effectiveness in establishing the low bit-rate operation and quality of coded speech. Here, the parameterization is based on the linear prediction filter model of human vocal tract wherein the linear prediction parameters (such as reflection coefficients, log-area-rations, line-spectral-frequencies, etc.) characterize the spectral shape of every frame of speech (typically of 20 ms duration) and the prediction error (residual) characterizes the excitation source of the speech production model. Coders have evolved to efficiently model and quantize both these information, i.e., the linear prediction parameters and the residual of each frame of the speech signal. However, by and large, the ultra low bit-rate coders have focused mainly on how to efficiently quantize the linear-prediction parameters either one frame at a time or by taking them as a sequence of frames together in the form of a ‘segment’. The main part of the book will focus on the various quantization schemes that have evolved towards quantizing the spectral parameters at the target bit-rates (of ultra low bit rate speech coding at less than few hundreds of bits/second). This chapter is organized to cover the various techniques in segment quantization, both conceptually and chronologically, as can be summarized under the following broad conceptual categories.