The Music Revival
Cedar Audio Restoration
As well as creating sound, digital technology can be used to rescue old or damaged recordings. Gordon Reid looks at a British system that's leading the world in audio restoration.
AS RECORDING STANDARDS RISE AND CDS BECOME CHEAPER AND MORE POPULAR, OLDER RECORDINGS ARE DECAYING. IF THEY ARE NOT TO BE LOST FOREVER, SOME METHOD OF MUSIC RESTORATION MUST BE FOUND.
FOUR YEARS AGO the British Library National Sound Archive decided to begin the transfer of their enormous collection of recordings to newer, more stable media.
Music stored on records and tapes deteriorates even if stored in ideal conditions. When a shellac record decays, the whole surface of the disk can break up and you're eventually left with nothing at all to play. Even in the early stages of decay, funguses eat into the surface causing significant surface noise and hiss. Tapes don't fare much better; the films become brittle, the oxides decay and, if badly stored, edges become damaged. And that's before you consider the action of moisture, and the flexing of the materials caused by changes in temperature and pressure. Clearly, you can only store disks or tapes for a limited time before they become useless - and for early 78s in particular, that time is fast approaching.
The advent of digital recording techniques (optical disks as well as CDs and DAT) meant that, for the first time, media were available that (in theory) were not going to deteriorate with time. The Archive therefore decided to investigate transcribing their collection onto this new media, and also to use the opportunity to "clean up" the material.
They purchased a Neve desk with sophisticated digital filtering facilities, and a "scratch reduce" facility - but this was unable to perform more than the most superficial cleaning up of material. The future lay elsewhere, so the NSA approached the signal processing laboratory at the University of Cambridge. The University convinced them that a computer-based solution was required although, in 1986, there were a number of technological problems still to be overcome.
Since all material would be processed by computer it would have to be converted to a digital format, processed, and finally played back as audio. The initial development of CEDAR (Computer Enhanced Digital Audio Restoration) was carried out using 12-bit A/D and D/A processors developed in Cambridge, but it wasn't until the advent of DAT that cheap 16-bit conversion and storage became available. The digitised material had to be loaded onto the hard disk of a computer, and an economical system was developed to take DAT (Or A/D converter) output and transfer this, via a special processor, onto the hard disk of a PC.
Finally, CEDAR was going to require phenomenal computing power - beyond the capabilities of even mini- and mainframe computers. This scale of computer installation would have made CEDAR unviable - who would pay the processing charges for a system running on a £1,000,000 computer? Early, and rather limited versions of CEDAR ran on 25Mhz 80386 PCs with 80387 co-processors. This configuration took a little over ten hours to process one side of a three-minute 78. It wasn't until February 1989 that a suitable microcomputer-based processor capable of handling over 50 million calculations per second became available. This enabled CEDAR to process a track in under one hour - still slow, but fast enough.
Combining these developments, CEDAR Audio were able to open for business. Initial processing speeds of 13:1 times real-time were soon reduced to 5:1 for some operations and CEDAR is now capable of processing both mono and stereo material, with a maximum track length of two hours. A two-hour track requires 1,800,000,000,000 individual calculations to process. That's over 250 million calculations per second of material.
CEDAR RESTORATION IS now performed by a subsidiary company based in Cambridge and imaginatively called Cambridge Sound Restoration. The most startling thing about their production office is the lack of heavyweight computing power. The only machines in evidence are Compaq and Dell PCs, plus a Macintosh network for company administration. The real computing power is hidden safely away inside the PCs. In addition to the computers are the banks of audio equipment - Sony DTC1000ES DAT recorders, reference quality amplifiers, and the inevitable Quad Electrostatic and Rogers monitors.
The specification of CEDAR is that it should be able to remove unwanted noise and hiss from a recording, but not interfere with the signal content in any way. Therefore it's not possible to use conventional filtering techniques, which will always act equally on signal and noise. Somehow CEDAR has to be able to differentiate between the two.
CEDAR Audio have grouped all the different manifestations of noise into four broad categories; first a recording is scanned for large disruptions such as gouges or even breakages in a record; secondly, the signal content is analysed to identify extended surface noise; next the clicks and scratches characteristic of record wear are located; and finally, the hiss (white or "coloured" noise) content of the recording is analysed. The first three of these categories are made up of individual events (fondly known as Snap, Crackle and Pop), but hiss is a continuous signal contained within the overall signal.
Consequently, CEDAR is constructed from four independent signal processing building blocks, and these are aimed at each of the specific classes of audio degradation. Within the four main processes are further sub-processes, which become progressively more specific to the elimination of a given type of degradation. This enables the CEDAR operator to apply a "sharp tool" to the degradation, rather than hitting the audio material with the proverbial sledgehammer.
THE LARGE THUMPS caused by deep gouges in a disc, the peeling away of fragile shellac surfaces, the seam of an early cylinder, or a break in a record, sound like large but discrete scratches. However, if they are treated by simple scratch removal techniques an annoying "pinging" sound remains. This is caused by the consequences of the thump (such as resonance in the record deck cartridge and arm) remaining after the initial event has been removed. Without the context of the thump these resonances are extremely distracting, and it may sound better to leave the recording unrestored. Consequently, both degradations must be removed simultaneously. Even with this system, a poorly performed restoration will still leave a low-frequency "shadow", and early Cambridge Sound Restoration work reflects this. Current versions of CEDAR can remove thumps in most musical material and, even in ideal listening conditions, the operation is undetectable. And since no editing takes place, the precise timing of the music or sound-track is always preserved.
THERE HAVE BEEN many attempts in the past to remove clicks and scratches, from old recordings. Both mechanical (reading the "cleaner" side of a groove) and electronic (using delay lines and impulsive limiters) methods have been employed, and these usually succeed in removing some of the unpleasant effects of scratches.
"MUSIC STORED ON RECORDS AND TAPES EVENTUALLY DETERIORATES EVEN IF STORED IN IDEAL CONDITIONS - FOR EARLY 78S THAT TIME IS FAST APPROACHING."
Unlike thumps, scratches are discrete events in a signal waveform. It would be relatively easy to remove all clicks from material if no genuine signal ever imitated a scratch. Unfortunately, there are many transient waveforms that, when viewed on a computer screen, look similar to scratches, but are nevertheless important constituents of the music.
To separate signal from noise, CEDAR scans the whole musical passage and then constructs a model of the signal what events are contained, and what reasonable boundary conditions can be chosen for the ranges of these events. (Examples of boundaries are maximum and minimum dynamic range, rise times and frequency responses). The actual signal can then be compared with the model and a decision made regarding unusual events. If an event is determined to be anomalous, it is then straightforward to eliminate it.
However, some scratches last for appreciable times (up to 15Oms) so there must be a mechanism to remove the scratch without causing a total break in the signal or resulting in time compression.
There are a number of ways to maintain a continuous signal after a chunk has been removed. These can be grouped Into three broad categories; splicing, re-synthesis, and interpolation.
Splicing, as its name suggests, involves cutting and pasting a section of signal into the gap caused by the scratch removal. The inserted section can be chosen from a number of sources and, if smoothed, will give a (mostly) glitch-free signal. However, there will often be be a loss of high frequencies (due to the smoothing) and, in any case, the method is ethically unsatisfactory.
Re-synthesis is a better, but more complicated method than splicing, and requires the construction of an artificial signal ("carved" from white noise) to fill the gap. This is real Honours Degree maths stuff, so we'll move on to...
Interpolation. This involves looking at either side of a gap in a signal and asking what signal "events" must have happened in the time for one to have become the other.
Simple in principle, interpolation is fiendishly difficult to implement because a musical signal doesn't follow simple mathematical rules. (Not with an 104-piece orchestra, natural reverb, EQ and studio enhancement it doesn't.)
CEDAR's approach to scratch removal is based on using each of these methods, or a combination of them, where most appropriate. To help their developers, they have devised a program which takes a known signal and removes a section from it digitally. The scratch removal process is run over this material, and the validity of the result can be tested against the original signal. The knowledge gained is then used to update the software. Clever stuff.
CEDAR has to deal with some very old and badly-degraded material. The maximum number of scratches yet encountered is over 2,200 in one second of music. The random nature of the clicks (in time, intensity and duration) creates a sound like bacon frying. (The term bacon" has become a quality control description - along with globular, glassy, frangy, and threatening.) As the number of scratches and their amplitudes increase, the bacon eventually dominates the output. In theory, there will eventually be very little signal left - other than the noise of the scratches. But even 2000 discrete clicks in a second of material can still leave up to 96% of the signal information intact, and CEDAR can handle this level of degradation very effectively.
THE CHALLENGES PRESENTED by hiss removal are some of the most important areas of current signal processing research. Unlike modern recordings, which can be made with very low noise, an aged recording cannot shed its hiss through conventional techniques. It's no good applying Dolby SR or dbx if the recorded material contains its own noise. CEDAR has to attack the noise at source. After all, if a relatively noise-free digital master can be made from the CEDAR restoration, subsequent copies of the material will only suffer from the noise associated with the modern playing medium - not from hiss recorded in 1940.
Noise is caused by the presence of random amounts of random frequencies in addition to the information contained within the musical signal. If high frequencies predominate, you'll hear hiss, if low frequencies dominate, you'll hear rumble. Unfortunately for the restorer, the noise is part of the continuous audio signal and cannot be differentiated in the way that scratches and clicks can. If a hissy signal is viewed on screen, the noise components are indistinguishable from the sound content.
The major problem for any restoration system, irrespective of the processing method employed, is therefore always the same: how to remove the noise without eating into the music. Compression will always occur if the signal is down-graded in any way, and yet it's a physical property of sound that whenever a signal subtraction occurs it is impossible to remove 100% of the noise, but leave 100% of the desired sound behind. Consequently, all previous attempts at noise removal have failed to limit the signal reduction to just the unwanted noise and thus suffer from serious amplitude and frequency compression.
"TO SEPARATE SIGNAL FROM NOISE, CEDAR FIRST SCANS THE WHOLE MUSICAL PASSAGE AND
THEN CONSTRUCTS A MODEL OF WHAT EVENTS ARE CONTAINED WITHIN THE SIGNAL."
One of the innovations of CEDAR is a unique method for determining the amplitude of noise frequencies at any time throughout a recording. Given that CEDAR can accurately track the changing noise characteristics, it can then adapt its reductive processes accordingly. Degradation such as surface noise (which can be inaudible one moment, and intolerable the next) can be tracked and the noise reduction dynamically tailored as appropriate. Since it's impossible to avoid removing some of the original signal during the noise removal procedure, the trick is to maximise noise reduction and minimise damage to the signal. Unfortunately, the actual mechanics of the subtractive processes are covered by worldwide patents and CEDAR are giving away no information regarding them.
CEDAR do, however, admit that the final result of processing can only be judged by listening. They've found that some clients require severe noise reduction, even if this leads to 3-4dB of compression, whilst others prefer to sacrifice noise reduction to ensure that there's no modification of the signal. Consequently, much time is spent matching the restoration parameters to given requirements.
Tracks may sometimes not be fully restorable because the quality of the input material is extremely poor. On these occasions you'd be offered partial restoration, and supplied with a number of differently processed samples. It's then up to you to decide whether any of the compromises are acceptable.
THE RANGE OF applications for a system such as CEDAR is impressive. It has now been applied to individual tracks and excerpts, complete CD mastering, international archive material, unpublished demo and practice tapes, and even samples for use in studios. But another area that's being rapidly revolutionised by digital technology is that of film and video. We're now seeing computer-generated graphics, both in new work and re-processed film - remember Fritz Lang's black & white classic Metropolis in pastels with Giorgio Moroder soundtrack? In a wider context, many old films are deteriorating rapidly and will be transferred to newer media to be preserved for future audiences. The optical transfer can be easily accomplished using modern video techniques, but up to now the sound has been on a "warts-n-all' basis. CEDAR can increase the value of newer prints by removing optical soundtrack problems and noise introduced by earlier copying processes. CSR presented a paper which was well received at this year's BKSTS (British Kinematic, Screen and Television Society) Show at Olympia, and a glance at the schedule in their main office shows work in progress for the National Film Archive, Thames TV and Channel 4. It can only be a short step to the first full-length, CEDAR-processed feature film.
The bureau in Cambridge is available for anyone to use - private collector, archivist, or record, TV, film or video company. Material is usually supplied on DAT but Cambridge Sound Restoration, via the National Sound Archive, currently offer one of the most accurate transcription services in the UK. The NSA can handle most forms of sound medium including 78s, cylinders, LPs, 45s, cassettes, spool tape, CDs, DAT and PCM formats.
Where possible, CSR prefer to work with a dry recording, free of EQ and other enhancements. If the processed material is scheduled for re-release these treatments can be applied after restoration to much greater effect. When people submit their material, they should ensure that Cambridge Sound Restoration receive the whole recording - including all lead-in noise and disc or tape run-out. This enables CEDAR to make the best use of all the information regarding the original recording conditions.
CEDAR's engineers feel that it's also important to discuss customers' specialist requirements. They can then tailor their restorations to the customers' exact specifications and they welcome customers' input at all stages of restoration. There are as many perceptions of "the right sound" as there are listeners and just to make everything as difficult as possible, UK record companies tend to aim for relative restoration (how much better is it than it was before processing) whilst foreign customers tend to impose an absolute measure of quality ("it's a fantastic improvement and you've obviously worked very hard, but it's still too noisy so we're not going to pay you...").
THE ACID TEST: how good is a CEDAR restoration? The answer has to be sometimes good, but sometimes fantastic. Cambridge Sound Restoration produce a demonstration DAT which highlights each of the restoration processes. To hear a broken, scratchy and hissy 78 transformed into a clear, bright recording is astounding. Other demonstrations on the DAT include noise removal on a damaged optical soundtrack, full restoration of a number of other 78s, and hiss removal on a selection of '40s and '50s masters. Each restoration is preceded by the original track and it's clear that there's no cheating going on. In fact, each demonstration is taken from work sent in by customers and many have been released, or are scheduled for release, on CD.
Faint modulation effects are audible on some restorations, and are more annoying on some than others. These are side effects of the hiss reduction processes and never occur on scratch or thump removal. Constant updating of the system is reducing the incidence of this all the time, and the record companies are clearly happy with the trade-off of high levels of noise for occasional "glassiness".
Although CEDAR has many further abilities based on sophisticated digital audio filtering and other manipulations, CSR make a point about not making judgements about how a recording should have sounded. The philosophy of CEDAR audio processing is to restore musical material to its original recorded quality. No attempt is made to compensate for effects such as wow and flutter which may be a consequence of the state of development of recording equipment available at the time of performance, or to add modern enhancements such as equalisation and artificial ambience. What CSR return to their clients is as accurate a re-creation of how the material actually sounded on the day it was recorded as it is currently possible to produce.
CEDAR Audio are currently developing additional processes to remove certain types of distortion as well as compensate for the frequency limitations of early gramophone recording systems. Early tests on the algorithms have been quite successful, but they say that it will be some time before these services will be available to customers.
Garbage In, Garbage Out is a common phrase in computing circles and it applies to audio restoration as much as to any other field of processing. Many recordings are transcribed poorly, record decks are poorly set up, tape machines have dirty heads, azimuth errors, poor speed stability... And these are the ones in professional studios. CEDAR certainly cannot make a bad recording good, but a good recording buried in noise can be improved to a degree that would have been unthinkable only a few years ago.
Current restoration charges are around £150 for a single track, and upwards of £1400 for a CD master. Compared to £100 per hour for the use of a mastering studio these costs are very reasonable, and within the reach of more affluent private customers, if not those of you convinced there's a No. 1 single in that old demo - if only you could hear it...
For the future, the possibility of a domestic CEDAR system is already being discussed. CEDAR Audio admit that the falling price of technology makes all things possible. However, the price of a single CEDAR processing system is currently about £13,000.
Although prices have dropped considerably over the last five years (you only have to look as far as the advances in digital reverb units) it'll be quite a time before £13,000 technology costs £130. In addition, this equipment doesn't process in real time, and CEDAR estimate that real-time hardware will cost well over £20,000 (to say nothing of the development costs). Until the computer industry can produce the technology that's required, and produce it cheaply enough to bring the price down by a factor of 100, we're not likely to see it built in on domestic equipment. Nevertheless, CEDAR Audio are optimistic that such a development is not too far away. They believe that real-time digital processing will become available in the next few years - firstly to record companies and production studios, then to the pro and semi-pro markets, and finally to the average punter. If such a system is going to be developed, CEDAR will certainly be at the forefront of the field. Unfortunately, too often has a world-beating device been a British development, only for the USA and Japan to overtake us after a few years. It's to be hoped that CEDAR Audio can stay the course. They are certainly a long way ahead of the field.
Since this article was written, CEDAR have achieved real-time processing with their noise reduction system. Simply play the damaged material into the system and, instantaneously, the restored sound plays through the speakers. The benefits of this are obvious, and bring the CEDAR "black box" much nearer. This innovation, amongst others, has earned Cambridge Sound Restoration a finalist's place at this year's British Computer Society Awards.
Feature by Gordon Reid
mu:zines is the result of thousands of hours of effort, and will require many thousands more going forward to reach our goals of getting all this content online.
If you value this resource, you can support this project - it really helps!