Audio Signal Processing and Coding

4.1 INTRODUCTION

Linear predictive coders are embedded in several telephony and multimedia standards [G.729] [G.723.1] [IS-893] [ISOI99]. Linear predictive coding (LPC) [Kroo95] is mostly used for source coding of speech signals and the dominant application of LPC is cellular telephony. Recently linear prediction (LP) analysis/synthesis has also been integrated in some of the wideband speech coding standards [G.722] [G.722.2] [Bess02] and in audio modeling [Iwak96] [Mori96] [Harm97a] [Harm97b] [Bola98] [ISOI00].

LP analysis/synthesis exploits the short- and long-term correlation to parameterize the signal in terms of a source-system representation. LP analysis can be open loop or closed loop. In closed-loop analysis, also called analysis-by-synthesis, the LP parameters are estimated by minimizing the “perceptually weighted” difference between the original and reconstructed signal. Speech coding standards use a perceptual weighting filter (PWF) to shape the quantization noise according to the masking properties of the human ear [Schr79] [Kroo95] [Sala98]. Although the PWF has been successful in speech coding, audio coding requires a more sophisticated strategy to exploit perceptual redundancies. To this end, several extensions [Bess02] [G.722.2] to the conventional LPC have been proposed. Hybrid transform/predictive coding techniques have also been employed for high-quality, low-bit-rate coding [Ramp98] [Ramp99] [Rong99] [ISOI99] [ISOI00]. Other LP methods that make use of perceptual ...