Goodness measure of speech reconstruction using the bootstrap

We address an application of the bootstrap for a goodness measure of speech reconstruction, such as a voice coder or noise reduction. The study is based on the spectral envelope of speech segments using autoregressive modeling. Employing the bootstrap, we are able to estimate the statistical distribution of regression parameters and therewith also reflection coefficients and the parametric spectrum. Confidence bounds are evaluated to calculate a segment-wise distance measure between input and output of the speech processing system.