A model order based sinusoid representation for audio signals

This work investigates the problem of sinusoidal parameter estimation including amplitude, frequency and phase in all kinds of audio signals including speech, music and mixtures using a nonlinear least-squares (NLS) estimator where the number of sinusoids is in general unknown. A solution to the sinusoid order selection problem is presented by employing the NLS method along with some information criteria (IC) such as minimum description length (MDL) or AIC. The simulation results demonstrate that the presented approach is successful in determining the number of sinusoids used in sinusoidal representation by using the information provided by well-known ICs´. It is demonstrated that the original audio signal can successfully be reconstructed in terms of both objective (SSNR) and subjective results (MOS).