The Detection and Correction of Artefacts in Degraded Gramophone Recordings

Tdep h o n e: (0223)-332 767 This paper presents recent developments in techniques for the restoration of audio material degraded by clicks. This type of degradation is associated with all forms of audio media, including CD and D.+IT, but is most characteristic of gramophone disks. The term 'click' covers a wide variety of problems ranging from loud isolated 'pops' to the relatively low level 'craclde' associated with most i s r p m records. The term 'scratch' is also used in this paper t o indicate the same type of degradation. JIost tj.pes of click can be modelled as bursts of corrupted audio samples occuring at random times and of randoni duration. For example, a poor quality isrpm record might tjyically have around 2.000 clicks per second of recorded material, with lengths ranging from less than 2Ops up t o 4ms. .A click remoi,al system is thus set the task of identifying the position and length of each indiyidual click and then replacing the click with new material in such a way that the listener is not aivare of any discontinuity. Rayner and I'aseghi designed a digital restoration system of this tj.pe in their \York a t Cambridge Unii-ersity before 19S9. The techniques used are model-based, assuming that a time-varying auto-regressive (AR) model applies to the audio signal. Detection is automated , identifying significant deviations from the current AR model as clicks. The position and length of each.click is then passed t o an interpolation algorithm. This minimizes the excitation energy over the gap, resulting in a linear least squares estimator for missing samples in terms of correct samples surrounding the gap. is now possible in real-time on modern DSP hardware. The system has been rigorously tested and developed t o such an extent that click removal One limitation of restoration performance is observed when the length of a click becomes large. Visual examination of waveforms shows that the restored signal often does not have enough energy towards the centre of the gap. For many audio signals the effect is visible for scratch lengths greater than 30 samples. Fortunately, the problem is generally not audible until much longer scratches are interpolated, greater than say 100 samples. This phenomenon is a major limiting factor on the maximum number of samples which the process may successfully interpolate.