LPC Quantization and Interpolation in Coding for Speech Storage Applications

In this paper, personalized quantization of the filter coefficients of Linear Predictive Coding (LPC) is studied. The study covers two aspects. On the one hand, a signal-adaptive algorithm which determines when to transmit a set of LPC coefficients is introduced. This algorithm allows a reduction of about 35% of the bit rate needed to code the LPC coefficients in speech storage applications – e.g. voice prompts in mobile terminals. The second part of this paper deals with the quantizing of LPC coefficients. Different approaches to quantization are compared in the context of speech coding for storage applications: Split Vector Quantization (SVQ), e.g. as used in the Adaptive Multi Rate Speech Codec (AMR) for GSM and UMTS, near optimum vector quantization with the LBG Algorithm (Linde, Buzo, Gray), and Lattice Quantization (LQ) with and without entropy coding.