There are an infinite number of processes that can affect the human perception of sound. For example, the live sound of an orchestra is dependent upon the venue, the audience and the local ambience. A recording of the same orchestra can be affected by a myriad more effects.
So what is the aim of a restoration engineer? The archival viewpoint suggests that such an engineer should present the listener with the most authentic reproduction of the original sound that can be obtained. But what about the creative influence of the recording engineer? With modern recordings, the ensemble sound often exists only on the recording medium, and many parts of it have probably never been through a microphone. Therefore the commercially minded engineer may, in contrast, attempt to generate a new recording more appropriate to its intended use. This use could be, for example, to please the public palate, or to represent accurately the sound of an era. Every restoration has its own criteria.
The algorithm designer is responsible for creating the facilities by which the restoration engineer generates new recordings from old. He or she does this by developing and implementing algorithms which remove unwanted sounds and/or effects present on the old recording.
CEDAR Audio provides a powerful set of restoration tools flexible enough to be used as the restoration engineer sees fit. However, it is our policy that the human ear should always be the final arbiter of sound quality; judgements based upon signal processing techniques are secondary considerations.
This section discusses some of the fundamental considerations used in designing a restoration algorithm.
Any sound recording has been through a process history; an example of such a history is shown in figure 1.
A degradation process can be described by the elements shown in Figure 2:
A restoration algorithm uses assumptions about the behaviour of these elements to restore the good signal from the corrupted signal, and the quality of the restoration depends upon the quality of the assumptions. If you can glean more information about a recording, you can devise a better restoration algorithm.
For example, if the uncorrupted signal is a violin solo, you may wish to include in your algorithm the sonic characteristics of a violin. Such assumptions should then enable the algorithm to differentiate between the effect of the degradation process and the good signal. These differences can then be used to regenerate the good signal from the corrupt. But problems will arise when your assumptions fail: i.e. when the assumptions represent an incomplete picture of the true degradation process, or do not accurately represent the good signal being restored. The assumptions therefore limit the number of recordings to which your algorithm can be applied. You should not expect your solo violin algorithm to work for a full orchestra or a solo pan-pipe.
A musical signal is random in nature, as are most degrading processes. Information theory tells us that the mixing of two random signals represents a loss of information, and that a perfect restoration is then impossible. A restoration algorithm therefore has to generate the 'most likely' good signal given the information available. Curiously, such an algorithm represents an additional loss of information about the degrading action - it has removed most of it. This has important implications for further reprocessing should a better algorithm become available: it is almost always better to work from the original recording rather than from an earlier processed version. A good example of this is found when removing the crackle found on 78rpm records: while the recorded signal may only have a bandwidth of 12kHz, the crackle will exhibit the full bandwidth of the reproduction equipment, so low pass filtering (which may offer a subjective improvement in signal quality) removes a lot of information that would be of good use to a more advanced restoration algorithm.
The final test for any algorithm is the human ear. The questions to ask are:
If the ear rejects the results as unsuitable for the intended application, then it will be necessary to redesign the algorithm.
The public has now accepted digital sound, and the debates regarding the pros and cons of analogue vs. digital have waned. However, there are other considerations to be taken into account when designing a sonic processor.
While a digital sound signal can be made into a near-perfect reproduction of a band-limited analogue signal, there are sonic processes that are native to each domain, and each can only approximate the other. For simple ideas, an analogue implementation is often the most cost-effective solution. However, most audio concepts requiring higher mathematics are impractical in the analogue domain. Digital Signal Processing technology has been developed to implement these higher mathematical applications.
The advent of computers and hard disks enabled a restoration algorithm to take as much time as it needed to fulfil its purpose. Unfortunately, this was a mixed blessing because Real-Time applications have the advantage of allowing direct feed-back between the engineer's ear and the algorithm's user-controlled parameters. How quickly a track can be restored also has important commercial considerations.
The order in which restoration processes are carried out makes a great deal of difference to the quality of the result. The correct sequence is declick, decrackle, debuzz, and then dehiss.
This is because large clicks make it difficult for the de-crackle process to identify and remove the tiny clicks and crackles that constitute surface noise, buzz, and other such problems. Furthermore, if clicks are presented to any dehiss process they confuse it and create unmusical side-effects. Conversely, dehissing first will make it almost impossible to identify and remove clicks and scratches at a later time.
Decrackling should be the second process because small crackles will also cause problems for the dehisser. Similarly, you should perform de-buzzing (when necessary) at an appropriate point in the total restoration process. This will generally be after de-clicking and often after de-crackling. Consequently, dehissing should always be the final process in the restoration chain.