Algorithms for the removal of scratches, impulsive disturbances, and white noise from archived gramophone recordings are presented. The case when there is more than one copy of a recording, with uncorrelated noise components, is also considered. The scratch removal algorithm is based on the observation that scratches are essentially the impulse response of the playback mechanism. Therefore it is expected that for a given playback device, various scratch pulses exhibit similar characteristics. The scratch removal system is composed of a matched filter-detector, a linear adaptive noise subtracter, and an interpolator for the replacement of irrevocably distorted samples. The impulsive-noise removal system is based on a noise-pulse detection-signal interpolation strategy. The noisy signal is processed to improve the detectability of noise pulses. Samples degraded by noise pulses are discarded and interpolated. The impulsive-noise suppression system consists of a linear-predictor-(LP)-based impulsive noise detector and an LP-based signal interpolator. The white-noise removal system is a variant of the classical spectral subtraction algorithm. However, a novel modification is included that largely overcomes the problem of musical noise and produces outputs with a greater degree of clarity and crispness. The modification is based on the concept of defining a time coherency (lifetime) for noise and signal components. A statistical decision as to whether a process is noise only, or delicate genuine signal structure is based on the signal-to-noise ratio and the signal lifetime. A method is described for the restoration of old recordings for the case when several copies of a recording with uncorrelated noise components are available. A major problem in using more than one copy of a recording is that the played back signals are not time aligned. Furthermore, due to small fluctuations in the speed of the playback device, the delay between signals varies slowly with time. An adaptive-noise-suppression system is used to implement a time-varying Wiener filter that time-aligns the signals and reduces the noise. The system achieves substantial improvement in signal-to-noise ratio. The suitability of least-mean-square, recursive least-square, and block least-square algorithms for implementation of the time-varying Wiener filter are considered.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.