Multi-sinusoid excitation model for audio coding

Abstract

Most studies in LPC-based audio coders decompose the signal into the product of excitation and system spectra, and then quantize the excitation by using either a stochastic codebook or multiple pulses. But their near white spectra cannot precisely describe the harmonic characteristics of excitation, especially when dealing with instrumental music. The paper explores the benefits of sinusoidal representation for excitation in the design of analysis-by-synthesis predictive coders. Furthermore, an efficient parameter extraction algorithm has also been developed to identify the associated parameters of the sinusoidal components. Simulation results indicate that the proposed multi-sinusoid excitation model allows the implementation of an LPC-based audio coder which delivers near toll quality at the rate of 92.61 kbps.<>