The performance of spectral quality measures

Two different classes of quality measures are discussed and compared: absolute and relative measures. The relative class, to which the prediction error belongs, has many different equivalent members, like the spectral distortion and the likelihood ratio. This type of measure is based on time series theory. The prediction error can be written either as a squared error of prediction in the time domain or as a relative error in the frequency domain. It is useful in many applications, especially in comparing models obtained with different estimation algorithms. It is compared to some measures that are absolute in the frequency domain. To that class belongs the integrated squared difference between spectra, that gives equal weights to all frequencies. Another measure is based on the squared difference between impulse responses. The absolute class has only a few practical applications, mainly in speech.