Criteria to measure the quality of TVAR estimation for audio signals

An audio signal can be represented by a Time-Varying Auto-Regressive (TVAR) model, whose parameters can be estimated by a particle filter. Since the original parameters are unavailable for real signals, an evaluation of the estimation may be traditionally performed through indirect criteria such as the SNR of the signal denoised by a Kalman filter based on the TVAR estimated model or through a statistical analysis based on the observation. We propose a new evaluation method based on the statistical characterization of the output of the inverse TVAR estimated model. The proposed criteria are much more suitable and coherent when correlated to the direct criterion (cepstral distance), which is related to the estimated TVAR parameters.