Gramophones have regained widespread popularity over the past few years. Being an analogue storage medium, gramophone records are subject to distortions which are mainly caused by scratches. This paper empirically analyses various outlier detection algorithms and proposes a novel predictive approach for noise detection. Twelve different forecasting models were utilized for the predictive deviation method. Once outliers are identified, they can be reconstructed using interpolation algorithms or time series approximation. Experiments were conducted on 800 songs from eight genres, both with artificial and real gramophone noise. The algorithms were compared according to their detection rate, computational speed and the tradeoff between accuracy and speed. It was found that the novel absolute predictive deviation using the autoregressive integrate moving average model performed best overall. The experiments also indicated that it was easier to detect noise in stable signals from genres, compared to noise in volatile signals.