In this paper, memory-based quantization is studied in detail. We propose a new framework, power series quantization (PSQ), for memory-based quantization. With linear spectral frequency (LSF) quantization as the application, several common memory-based quantization methods (FSVQ, predictive VQ, VPQ, safety-net, etc.) are analyzed and compared with the proposed method, and it is shown that the proposed method performs better than all other tested methods. The proposed PSQ method is fully general, in that it can simulate all other memory-based quantizers if it is allowed unlimited complexity

Many current speech coders use a source-filter model in which the signal is decomposed on a frame-wise basis into an excitation component and a filter component. The filter component is then parameterized and coded by a separate spectrum encoder. The performance of such spectrum encoders is often evaluated in terms of a spectral distortion measure that treats each frame independently. Doing this therefore neglects the time evolution of quantization errors. In this paper, we focus on the time evolution of spectrum vectors in speech spectrum coding. We show that the human auditory system prefers a slowly evolving spectrum quantization noise, resulting in smooth spectrum trajectories. This result is the basis for a new distortion criterion for spectrum coding which introduces interframe memory into the standard spectral distortion measure, thereby having the ability to penalize a poor spectrum evolution. Further, a number of coding schemes for explicit spectrum evolution control are evaluated, and we show gains in the order of 2-4 bits, compared to schemes with no evolution control.