Various problems cropped up during the restoration of the soundtrack of the feature film The Mistress of Atlantis which are typical of those arising from the bad condition of old films. Attemps to eliminate the damage occasionally tested the limits of existing processing systems, which I would briefly like to summarise below.

Automatic Prewashing

The importance of a good scan as a prerequisite for further sound processing has been recognised at least since the Berlin FIAF-Congress 1987: "Archiving the Audiovisual Heritage". Wet scans and the narrowing of the light beams (in order where possible to skip surface imperfections) are techniques which were described at the time. This notwithstanding, I would like to emphasise that one should where possible always take care to use an earlier generation, otherwise every otherwise avoidable fault will itself be recorded on the soundtrack during the re-recording, thus making it impossible to extract it from later scans.

In order to keep their options open for future, semi-automated restoration as described below, I would recommend that archives consider the production of one-to-one prints, without filtering or cutting out high and low frequencies. If high frequencies are cut out, information valuable when declicking and decrackling is also being lost. It also means that any automised process will be made more difficult or indeed impossible, depending on the intensity of previous filtering. Great care should therefore be exercised in a process which at first appears instinctive but can in fact lead to faulty "improvement" of a print! This is even more important if one considers that the (nitrate) original is usually destroyed or disintegrates and is therefore lost forever.

For my work I use Sonic Solutions because, being a hard disk system, it offers a manual editing facility, which has proved to be an important tool. The first stages of the process involve the mainly automatic declicking and decracking programmes. Unfortunately, each of these processes takes from eight to ten times as long as the actual sound material, although the manufacturer has announced that future versions will feature real-time processing.

A click is a short, needle-shaped, wide band disturbance peak which in the case of Figure 1 is followed by a low frequency thump. The system searches through the frequency spectrum to discover the characteristics of this peak, which we hear as a click, then encloses this impulse and deletes the signal. The signal before and after the click is used as a guide to reconstruct the content at the deleted spot. In this sense, the signal is not really reconstructed, since there is no way of knowing what it looked like: instead, one pretends that the transition from the signal and after the click has been briefly interrupted. Since this hiatus is so brief, the result in most cases is good and no artifacts (i.e. new disturbances replacing the old ones) appear. Since hard disk systems do not normally write over an original but add new data to the original sound information, this enables the user to reconstruct the original signal at any time should the substitute signal not be satisfactory. The advantage of such hard disk management is that the user can correct each disturbance directly until such time as a satisfactory result has been attained.

Crackles are comparable to the surface noise on old records. Unlike the above-mentioned clicks, the crackle is a constant accumulation of tiny clicks. Like clicking, the computer recognises crackling by means of its frequency spectrum, although tiny crackles are much shorter and numerous than the rougher clicks previously removed in declicking. The decrackling gives excellent results, especially for old films.

Manual Interpolation

The types of disturbance which do not fit into the limited scheme of automatic fault finding must be removed manually. Whenever a click-like disturbance appears this can be done with the aid of interpolation. The click can be either added to a click list, so that it will be interpolated during the next automatic run-through or it can be gated by putting a cursor to the left and right of the click, then giving the command to interpolate, i.e. doing manually what the computer would do automatically. With this method it is possible to influence directly the process by which the signal is substituted. On the one hand, it is possible to choose fairly freely the length of the section to be substituted, which means that long sections can be substituted provided this seems useful; on the other hand, it is possible to choose the side from which to approach the section for interpolation. In this way, I could, for example, substitute a relatively short, damaged passage from the first note of a loud piece of brass music by deciding to interpolate from the right rather than the left; otherwise the section prior to the music would have been used as a guide track, the beginning of the music would be lost and the fanfare would begin too late.

I also used interpolation for restoration in the section "took me to hospital" (see also Figure 2). The word "me" had been chopped up and two parts of this word's signal were missing. Although these missing sections were very short, their absence was noticeable. Interpolation here made it possible to refill the word with sound and restore it to its original length. This time I enclosed the short passage lacking a signal with both gates, interpolated and was thus able to reconstitute the missing sections.

Jumps in Sound Level

Abrupt jumps in sound level were a regular problem, particularly during quiet passages. These jumps in sound levels varied greatly in length; there were extremely short peaks which lasted no longer than half a frame to one frame as well as longer passages of a higher sound level lasting several seconds. Using a hard disk editing facility, it is not especially difficult to remove jumps in sound level. Nevertheless, the sound engineer's skills are sometimes put to the test, especially if very short fluctuations are to be removed, making them inaudible. Sonic Solutions provide two methods of dealing with this kind of problem: the damaged sections can be enclosed with two crossfades and the volume between the two fades can either be raised to up to 6 dB or lowered to zero, or, a second method which allows volume transitions to be matched precisely, is to mask the section with a single crossfade. A crossfade consists of a fade-in and fade-out that can be adjusted according to certain trigonometric functions such as cosinus, arctan and others.

One particularly delightful challenge was a section of dialogue: "They took me to hospital" at the beginning of reel 5 (see Figure 3). There was an inexplicable drop in sound level of approx. 15 dB which the system was not able to level out in one run-through. The mixing desk is able to raise the level by more than 6 dB, but it is not possible to control this with the necessary precision. The drop in sound level can be repaired initially if one re-records the section at a suitably higher level (up to 20 dB) into a new file using the mixing desk, subsequently adding this to the original sound. Unfortunately, since the signal is so weak, we have now also increased the naturally strong noise disturbance and must therefore denoise this same "took" separately. A trace of this hiss can still be heard, but I believe that the dialogue is now at least comprehensible from beginning to end and one is no longer shocked by the brief disappearance of the signal.

In roughly half the cases the short peaks were part of a low frequency disturbance which could not be masked by simple level adjustment. These often very loud low frequency disturbances pose a particularly difficult problem for the restorer not only on account of their length, which makes it impossible to just delete them and replace them by interpolation. (Although interpolation would work, the result in these cases is rather poor and terrible artefacts can be heard). Their spectral characteristics are also problematic since they often contain a considerable amount of high frequencies and it is not enough to filter out low frequency sections or even individual frequencies directly.

What can be done about this? Unfortunately, Sonic Solutions does not offer a feature which can automatically deal with this problem. Here methods must be adapted to the specific characteristics of the signal. If we are dealing with a case of simple noise, then the problem is easily solved: the damaged noise is simply replaced by intact noise, rather like working with a word processor. However, should we find ourselves in the midst of a piece of music or another very complex form of signal, it is sometimes only possible to filter out the low frequency section which causes the most disturbance. This can, however, be timed fairly precisely so that the signal preceding and following the disturbance remains unfiltered. In general all filtering and mixing carried out at the mixing desk is fairly accurate if programmed in advance. The editing mode, including crossfade, manual interpolation and so on, allows accuracy to within one thousand of a frame, which is sufficient.

Another delicacy was the removal of twenty or thirty intermittent disturbances during a passage of music. This passage of music sometimes repeated itself, so that, after a lengthy search it was possible to replace damaged sections. The result is surprising: nothing recalls the low frequency thumps that had destroyed the oriental atmosphere of the music. Fortunately, one does not have to keep an eye on synchronicity, since Sonic Solutions contains a "three point editing facility" which maintains the synchronicity of the original.

Denoising

The problems of denoising technology have not yet been satisfactorily resolved, which is a real problem for anyone attempting to restore old soundtracks. The setting of the parameters is critical, as is the assumption, basically determined by the system, that noise is in principle static - i.e. its volume does not really change (within a particular section). Whilst one still has to filter within sharp defined parameters in order to extract from the working signal exactly that spectrum within which its level is to be disregarded, one should not set the parameters which control the process, especially in the case of old, noisy copies of films of the 1930s, too sharply due to the fact that one can otherwise very quickly totally mess up a soundtrack through incorrect filtering. Depending on the level of phase modulation, this can then sound worse than the original noisy recording, and not only to purists' ears. For this reason, the effectiveness of denoising is limited, particularly in the case of the very films that need it most, because the levels of the noise and the working signals are very close together and the signals can only be separated by very filtering within very sharply defined parameters. In this case, a little means a lot. Because we cannot reckon with a process in the foreseeable future which will enable copies of the original to be reconstructed from the available recording, the subjective assessment - i.e. using our own ears to make a comparison between the original and the copy - remains the only standard for ensuring a successful treatment.

Conclusion

Editing takes up the most time because the removal of many faults cannot be automated. In addition to this, individual decisions are often complicated and require a feel for the work. Thus time costs money and in the future will be at best only partially alleviated by automated processes. The development of software which detects and rectifies at least low frequency disturbance is conceivable, as is the development of improved denoising software, already nearly ready, and of scanners which first read a soundtrack as graphics which they then digitally process and only then turn into sound. Thus they provide better preconditions for an extensive treatment of original soundtracks - insofar as these originals still exist. However, the fully automated restoration of films like the Mistress is still inconceivable to me as long as one wishes to remain faithful to the original version.

Klaus-Peter Schmitt

This research was carried out in the frame of the NIPKOW program, Berlin 1992-93.

Klaus-Peter Schmitt a résumé les difficutlés rencontrées. Le rapport de Schmitt contient des commentiaries précieux sur les problèmes qui se posent lorsqu'on souhaite éliminer des cliquettements, des grésillements, des sons de basse fréquence et des sifflements dans la bande sonore de duplication de copies anciennes (le choix d'une bande de génération aussi jeune que possible est, par ailleurs, d'une grande importance).

L'étude comporte des observations effectuées lors des divers procédés intervenant lors de duplications à des fins de préservation, tels que la détection automatique des altérations, l'interpolation manuelle, le traitement des modifications abruptes dans le volume sonore ou des perturbations de basse fréquence, la technologie de déparasitage, etc.

Le développement de software permettant de détecter et de rectifier les perturbations de basse fréquence ainsi que celui utilisé pour le déparasitage est concevable. De même, nous disposons déjà de scanners lisant la bande son comme un graphique permettant ainsi d'éliminer certaines imperfections. La restauration totalement automatisée d'un film comme l'Atlantide, reste cependant inconcevable dans l'état actuel de la technologie.

Clic, crac, bum y silbidos. Experiencias con la restauración de bandas de sonido

La restauración de la "Atlantida" (1932) de G.W. Pabst permitió examinar los límites de los procedimientos utilizados corrientemente en la restauración de las bandas de sonido de películas antiguas.

En su detallado informe, Klaus-Peter Schmitt, aborda los problemas que se plantean cuando se trata de eliminar clics, chisporroteos, estallidos de sonidos de baja frecuencia y silbidos que aparecen durante la duplicación de material antiguo.

El estudio de Schmitt contiene observaciones efectuadas durante los procedimientos de preservación tales como la detección automática de alteraciones, la interpolación manual en la reconstitución de la pista sonora, el tratamiento de saltos bruscos en el volumen o de perturbaciones de baja frecuencia, la tecnología de deparasitaje, etc.

Un software que permita detectar y rectificar perturbaciones de baja frecuencia o eliminar sonidos parásitos está en buenas vías de desarrollo. Asimismo se díspone ya de scanners que analizan la banda de sonido gráficamente, permitiendo de esta manera su reconstitución. La restauración totalmente automatizada de la banda de sonído de una película como la "Atlántida" es aún inconcebible con los medios de la tecnología actual.