An algorithm for the correction of disturbances or gaps of up to several thousand samples in an audio signal is presented. The reconstruction is based on a novel method for time-domain discrete signal extrapolation. The missing or disturbed portion of the audio signal is replaced by a weighted average of signals extrapolated from the areas preceding and following the disturbed portion. Impulsive-type errors usually distort the underlying signal irreversibly, and the damaged signal portion does not contain any information of the original signal. In the proposed method the damaged signal samples are not used in computing the replacing samples. The reconstruction method is applied in practice to correct scratches from signals recorded from badly damaged vinyl recordings. The proposed signal reconstruction method can be implemented in real-time applications.