You are here

MARVIN CAESAR: The Aphex Approach To Compression

Interview | Manufacturer By Paul White
Published January 1997

Aphex is celebrating 21 years in business, during which time the company has diversified from Aural Exciters into signal processing such as compression. However, Aphex President Marvin Caesar feels that the subject of compression is still widely misunderstood. Paul White asked him to set the record straight.

Aphex first rose to fame with their legendary Aural Exciter, first commercially marketed back in 1975, but since then, the company has diversified into signal processors, many of which handle some aspect of dynamics control. In a market crowded with 'me‑too' compressors and gates, I asked Aphex President Marvin Caesar how the company went about designing products that would stand out from the crowd.

"Our philosophy is that we should make unique products — and that came from my laziness, because I didn't want to have to compete with anybody! I didn't want to sell on price or just the odd feature, I wanted to make something that was a little bit better. The problem is that as the market has expanded downwards, there are fewer people who really know what they're listening for, and as a result, our equipment has to be explained to more and more people. And that gets me into gain reduction — most people don't fully understand what a compressor is supposed to do."

Was compression the first area you went into after the Aural Exciter?

"Yes. Most compressors are effects — the attitude is that if you put something through a compressor, you get a certain sound. When tuned properly, compressors can produce a very pleasant effect as you hear the artifacts of the gain control. Our philosophy is: fine, we'll give you the tools to do that, but the real trick of gain reduction is to be able to do it without introducing any effects other than level correction. In that regard, we've succeeded marvellously, first with the Compellor for average level control, then with the Dominator for peak level control. Our Easyrider has an intelligent compressor that is easy to use, but achieving that goal has confused a lot of people. A lot of people listen to our processor and say 'Why do I need this? I can't hear it working'."

This might be a good point to talk about the situations in which you might want an invisible gain control rather than compression as an effect.

"One important area is background vocals, which have a tendency to come up to the level of the lead vocalist or disappear. In a mix session, you can ride the faders to keep them 'in the pocket', but a good compressor will do a better job because you want to have them sit in a very narrow dynamic range. But you don't want it to sound in any way processed, you just want it to sit.

"Another example is where you have a music bed that you're trying to lay into a video, where the music is quite dynamic, but you want to keep it underneath the dialogue. Again, you need a transparent compressor, otherwise you hear the tonality changing as the compressor operates. Good compressors are also important in broadcast and PA, where you have to get above the noise floor, but you can't exceed a maximum level. Certainly, in terms of reducing the dynamic range of a 24‑bit digital programme down to VHS cassette, with its restricted dynamic range, requires a very transparent compressor."

So your aim has been to allow the user to apply as large a reduction in dynamic range as possible without the subjective sound of the material being affected. Perhaps this is a good time to look at how you do this, because on the face of it, all you have to play with are threshold ratios, attack times and release times.

"Before you even get to the intelligence of the side chain, you have to look at the signal path. Our philosophy has been that every processor must be able to do nothing before it does something — which means that it has to be absolutely transparent when you put it in line. The quality of the audio must not be changed by the fact that you're putting it through the processor. This requires a good choice of integrated circuits, and a good choice of input and output stages that interface transparently with anything you might want to plug into."

This is a problem area with many balanced units, because when you plug them into an unbalanced insert point, you lose 6dB of level.

"What we use is servo‑balanced inputs, as well as servo‑balanced outputs. The idea is that when you ground one leg of an output, you get the full signal swing on the other leg so you don't lose any level, and a lot of modern equipment works this way, but the problem with most of these circuits is that the grounded output stresses the output amplifier, and you're pumping current onto the ground. Our outputs work so that when one leg is grounded, you get the full voltage swing on the other leg, but we turn off the grounded leg.

"The real difference is that we use our own 1001 VCA. Many VCAs sound and perform quite well when they're not being exercised, meaning that there's no changing control voltage on them. But of course the whole purpose of a VCA is that you apply a control voltage to change the gain, and one weak area is the change in DC offset that occurs as you exercise the VCA. You hear that as clicks or pops in a noise gate; in a limiter you would hear it as the limiter working harder than it needs to, because the DC component at the output is being fed into the detector circuit. When there's no DC on the output — or no control voltage feed‑through, as some people call it — the processing is a lot more accurate. And of course the quality of the audio passing through is extremely good.

To make a compressor sound extreme, apply 10db of gain reduction in manual, hard‑knee mode, with a high ratio.

"The first test you should do on any compressor is to do a hard bypass and listen carefully to the sound to see if it changes in any way from passing through the processor. A kick drum or something with a very hard leading edge is probably the most revealing — look for a loss of that leading edge, a smearing of the high end or woolliness, and a loss of bass presence. Even though the device may measure perfectly flat, there may be something dynamic that's going on causing a loss of bass presence.

"Why is our circuit better? In the Expressor, we have a fully manual compressor, where you can adjust the release and ratio — all the parameters you find on a compressor — and indeed, you could adjust the parameters for each type of input you might ever look at, and if you're really good, you could adjust those parameters in real time, so that as the texture of the track changes, the compressor action changes with it. Of course, that's impossible in practice. There has to be some intelligence that allows the processor to constantly fine‑tune its own performance."

Aspects Of Compression

"When setting up a compressor, what you hear is that the higher the ratio, the faster the attack, and sometimes the release, the more the effect on the audio quality. If you try to achieve a better sound by slowing up the attack, and lightening up on the ratio, then you have the probability that the level will overshoot. Because of this, you have to further lower the threshold, which means now you're doing more processing, or you have to speed up the attack time, which still means more processing. What you want to achieve is the impression of a slow attack, because what we very often listen to are the leading edges of a signal. These relate to intelligibility — you recognise the consonants and interpolate the vowels — even if you don't hear vowels, you can still have intelligibility."

Are you saying that a slow attack and slow release will produce the most natural sound, but then the compressor doesn't get a chance to do its job properly? If I was presented with this problem, I'd hazard a guess that what you need is a dual‑stage processor, where a slow attack and release compressor is teamed with some kind of fast‑acting limiter to take care of the overshoots.

"That would depend on what type of processor you're talking about. There are different approaches that we use, depending on the application, and one of the ideas is that it's always better to run the signal through as few stages as possible, because no processing always sounds more natural than processing. There are so‑called compressor/limiters, and most compressor/limiters operate so that at the onset of gain reduction they're compressors, and as you get deeper into gain reduction, the ratio increases towards limiting. Soft‑knee type compressors fall into this category. Then there are other devices, where you have separate compressor and limiter sections, and they can work effectively, but the two sections must communicate with each other, otherwise you can run into trouble, depending on the texture of the input."

The other type I'm familiar with is the variable‑ratio compressor, where above around a 10:1 ratio, you can say that the compressor becomes a limiter. However, unless you have a very fast attack time, some transients will still get through before the compressor has time to react, so these can't be true limiters.

"Exactly. And one of the problems is using the right process for the application. You can use clipping, which is fast, infinite‑ratio limiting, but clipping is different from gain reduction. With gain reduction, the entire waveform is squeezed, whereas with clipping, you're taking the top off the waveform. Because of this, a gain‑reduction device is bound to sound less loud than a device that uses clipping."

Of course, the instinctive reaction is to say that you can't use clipping, because clipping is a particularly nasty and audible form of distortion.

"Clipping is audible if it is allowed to exceed a certain duration, and that duration is decreased if the clipping is repeated. The first clip can be a little bit longer, and then the subsequent clips need to be shorter, but when done properly, clipping does sound better, because it's actually less damaging to the sound and you don't lose the impact of the leading edge."

You must be talking about clipping of such short duration as to be inaudible. Is this in the region of 1ms or so?

"Actually, the first clip can be a little longer than that — you can get away with up to around 5ms first time around. However, if you look at a lot of mixes, you can see that the bass waveforms are flat‑topped, and that's not because the whole mix was clipped, but because the bass track was clipped before mixing. If you then put clipping on the peak excursion of the complete mix, it can make it sound worse.

Our philosophy was that we should make unique products, because I didn't want to have to compete with anyone!

"The idea is to know what you're trying to do to the signal: do you want to bring low level signals up, do you want to keep low level signals down, do you want to change the sound by deliberately introducing pumping and breathing — which can be quite musical — or do you want to both control level and introduce pumping and breathing?"

Virtually anyone can build a compressor that will pump and breathe, so what's the secret to making one that is as transparent as possible?

"Understanding the physics of sound waves and psychoacoustics. For example, an application of psychoacoustics is exploiting the fact that the human hearing system can't detect very short periods of clipping. Another example is in the Easyrider, where different attack times allow the initial change above threshold to go through with a slow attack, then if it exceeds a certain level, it engages a faster attack, but because the initial part of the transient went through on a slow attack, the impression is that the whole signal was treated with a slow attack. The result is a cleaner sound quality than if you use a second device to handle the peaks."


Perhaps we should look at where we should use compressors and where we should use limiters, because inexperienced engineers seem to interchange the two terms.

"If you use a manual compressor, when you set a high ratio and fast time constants, you'll hear that they have a major effect on the audio — loss of transients, loss of high end, the kick drum making the vocalist 'duck' [not physically, but in the sense of reducing the level of the vocal during kick drum hits — Ed], and so on. We call this spectral gain intermodulation. Once you've tried this and understood it for yourself, then you'll understand that if you want to bring low signals up, you have to set a low threshold. But if you have a low threshold, an awful lot of your signal is going to be above threshold, so whatever happens to that signal needs to be fairly gentle in terms of ratio and attack and release characteristics.

"If, on the other hand, I want to keep high‑level signals down, to do that effectively I have to have a fast attack time and a high ratio. As already stated, that has the maximum effect on the sound, so I don't want to put my entire signal above that threshold — the threshold must be set higher. Understanding what you want to do is critical in terms of where that threshold is set. With our compressors, the type of compression tends to be very gentle, but if you use the Expressor in a manual mode, set a high ratio and fast attack, it will behave like a regular dumb compressor — you'll hear it working. It sounds most transparent in auto mode, where the attack and release constants track the programme dynamics.

"If you look at a product like the Compellor, there are three separate brains that are always adjusting the attack and release. They're always analysing the relative density of the input signal. One of the biggest applications of the Compellor is for cable TV. Everyone has experienced how the commercials wake you up at night because they're so much louder than the programme, but you'll find that both the commercials and the programme have exactly the same peak output level. It's the average level that's different because the commercials are already heavily processed. The Compellor reads that and pulls up the release time on heavily‑processed signal and speeds it up on unprocessed signal. It does that totally automatically, so the segues between the two are much smoother."

On the other hand, you can sell people processors that will enable them to create very loud‑sounding commercials.

"That's the Dominator, and in terms of perceived loudness, it's interesting to note that if a sine wave produces 100% loudness, a square wave of the same peak level will register 158%. Now electronics respond to peak levels, especially digital equipment, and if you flat‑top a waveform, it's clearly going to be louder than reducing the gain."

It's all down to area under the waveform, and the maximum possible area is produced by a square wave.

"That's right, and here's something that's very interesting. Take a waveform with a high transient peak: if this is reduced using gain reduction, the whole waveform would be pushed down in level, even though the peak might be very short in duration. What's more, if the release time is long enough, whatever follows will also be reduced in gain.

Understanding what you want to do to a signal is critical in terms of where you set your threshold.

"If, instead, we clip the brief transient, the rest of the waveform isn't subjected to gain reduction, so the perceived result is much louder. And because the clipping is very short in duration, it's completely inaudible. Intelligent clipping is why the Dominator makes things sound so loud. The specific invention within the Dominator is the automatic limit threshold, where we monitor how long the signal has been in clip, and if it's been in clip too long, we reduce the threshold of the band limiters underneath the clipper. Nobody else does this — it's our patented process. We can set the threshold right at the maximum level for the following device, for example, to use all the bits in a digital recorder or to get to 100% modulation in a broadcast transmitter.

"Once you go into the digital domain, there's no way you can buy more resolution. You can cover garbage up with noise, but you can't get more resolution. Even with a 24‑bit system, you want to use up all your bits to keep as far away from the noise floor as possible. Also, there's a lot of data compression used in broadcast and storage, so to ensure that these have minimum impact on the audio quality, it's once again essential that you record at the maximum resolution possible. What's more, it helps if you can reduce the signal entropy — by entropy we mean anything unusual that's making the bit‑rate reduction system work harder. For example, if you can reduce transients before you go into a data compression system, the system is much better able to deal with it — the effects of data compression will be more benign.

"Using the Dominator is fairly benign unless you limit too deeply, and in my experience, the first 3‑6dB of limiting is free — you can't hear it. In fact, the Dominator is the limiter of choice for use with professional in‑ear monitoring systems."

Split Band

Do you have any views on the benefits or otherwise of split‑band compression, where different sections of the audio spectrum are compressed separately?

"When you're doing a wide‑band compressor, everything is being reduced, which is why the intelligent side‑chain control is needed to minimise it. The problem is that if you have something dominant in one part of the spectrum, it will tend to affect everything else.

"With split‑band devices, the effect you hear most is at high ratios — which is more like limiting — and with limiting, the threshold should be set high. Split‑band limiters make a lot of sense, but what some people do in broadcast is to use split‑band compression, which essentially re‑equalises every song. If you don't want to change the EQ, wide‑band compression and levelling is better, if it's done properly, whereas the purpose of multi‑band compression is to change the EQ.

"Our only current multi‑band unit is the Dominator limiter, which has three bands, but our next FM processor will be a multi‑band compressor, simply because the radio market demands a very loud, thick signal. There are lots of programme directors out there who think they know better how to mix a record than the artist or the producer, and that's an unfortunate situation, but I have some recommendations to producers on how to package audio so that broadcast processors do not do as much damage as they do to unpackaged audio."

Turning now to the possibility of digital dynamic controllers, is there any advantage to having a digital processor that can look a few samples ahead, so that it can anticipate what's coming next?

"Not really. In theory, it's always helpful to see what's coming, but given the attack and release characteristics of a good compressor, it's better for the side‑chain to look at the output than it is the input. And if you look at the output, it already happened. As far as I can tell, the existing technology is as good or better sounding than anything that can be done digitally."

Practical Tests

"To make a compressor sound extreme, apply 10dB of gain reduction in manual, hard‑knee mode, with a high ratio. Speed up the attack time and use a fast release time — what you should hear on a full mix is a loss of transients, and the bass modulating the mix so that it sounds distorted.

"Next, slow up the release time. You should hear some of the distortion go away, but the loudness will reduce.

"The other experiment is to reduce the attack time, and you'll hear the transients come back up as you turn the knob. If you have peak‑reading meters on the equipment following the compressor, you should also notice the peak levels coming up dramatically as the shorter transients pass through the compressor unchecked.

"Another thing to try is reducing the ratio, and here you'll notice a lot less spectral gain modulation — the bright sounds won't be so badly affected by what's going on at the bass end. By doing these experiments, you'll get a feel for what the controls actually do. For example, if you have higher ratios and very slow attack and release times, then you have levelling. And with a digital recorder, it may be beneficial to have some kind of leveller followed by a processor that limits the peak level of the signal — such as the Dominator.

Hard Or Soft Knee?

"The idea of a hard knee is that sometimes you want to hear the effect of the sound going over threshold — a sudden 'grab'. Soft knee allows a much smoother transition into the compression, but at the same time, you're doing a lot more processing before the ratio that you've set. In fact, what we do in the Expressor is to lower the threshold by 10dB, so you get approximately the same output when you switch from soft knee to hard knee."