Something that has taken me far too long to appreciate, is the wisdom behind the way SSC scales the output gain of their Kyma Compressors. With unity output gain, these act just like a compressor with automatic makeup gain, but that gain is such as to limit the maximum output level to around -10 dB. A maximum level signal into the Compressor would result in an output signal of -10 dBFS.
Somewhere, SSC does state that they view the -10 dB level as the nominal 0 dBVU level. I know I have seen that. But I spent far too long trying to overcome this margin provided by SSC, in an effort to get my signals back up to the input levels.
WRONG!!
What happens during a transient peak going through a compressor? The compressor at that moment has some additional gain applied, over and above unity gain. Compressors don't generally act instantaneously, although you can set an attack time of 0 in the Kyma compressors. But with any non-zero attack time, it takes a while for the compressor to realize that it must change the gain. Meanwhile the transient makes its way through the compressor with that initial higher gain applied.
During the time that the transient sneaks through, you can view the compressor as a linear device -- nothing more than a gain block. So a 6 dB rise in the transient level will produce a 6 dB rise in the output level as well. Not 6 dB divided by the compression ratio. That ratio only tells you what the eventual steady-state gain of the compressor will be, but until it can reorient itself to a stronger steady signal, it actually applies a unit slope to that signal, on top of the extra gain already provided by the compressor in its pre-transient state.
Hence, by allowing an output margin of 10 dB, SSC has allowed your transients to be as high as 10 dB or more than the rest of the signal without going into output limiting or clipping.
So the lesson here is that one ought not to try to work the compressor post-gain too hard, lest transients actually clip. The only time it is safe to provide the remaining 10 dB of makeup gain is when you are using an attack time of 0, but that generally squashes the sound so much that it doesn't sound so good.
Headroom, headroom, headroom....
--
DavidMcClain - 16 Nov 2004
Okay now with that understanding in place let's look at some improvements...
It should be clear that with the output limited to -10 dB for a full scale input signal, it is not possible to have any transients at that loud limit that could possibly push up above -10 dBFS on output.
However, with high compression ratios, it is concievable that lower level steady state signal levels could have high compression gain and a transient could push through that will go above that upper limit. -10 dBFS is probably a bit stingy here for most reasonable compression ratio, by about 4 to 6 dB.
But instead of liberalizing that limit, here's another idea that is even better....
Suppose instead of fixing an upper output limit of -10 dBFS, we relax that to allow full scale output for full scale input, and place an upper threshold on compression at -10 dBFS. Then signals above that level will not be compressed at all and will receive unity gain. We end up with a 2-breakpoint compression curve with the upper threshold fixed at -10 dBFS.
This new compression curve more resembles the "vintage" compressors that everyone raves about -- and for good reason. Transients on faint signals make their way through the system with minor overshoot (not damaging enough to overload the output stage), and strong steady-state signals that don't need any compression go through unchanged. Strong signals probably impy strong transients -- and these are also uncompressed and so cannot overshoot.
The amount of overshoot from a compressor is equal to the gain reduction that would be needed at that peak level. So if a compressor is running with a 2:1 ratio and a signal is given 6 dB of boost at its nominal level, then a 6 dB transient will produce an effective overshoot of 3 dB. That 6 dB increase, if it were steady state, would cause the compressor to drop its gain by 3 dB. Since the signal rises by 6 dB, it is overshooting the mark by 3 dB.
For signals above -10 dBFS, there can be no overshoot since the gain is unity and the compression ratio is also 1:1. You can't develop more output level than what you feed into the compressor, when the signals are in this upper region. The only signals that can produce compressor overshoot are those below -10 dBFS steady-state and with a transient.
We can take the standard Kyma Compressor and produce a version that works this way... It takes two compressors in series to do this.
There is another point to examine here... That of spectral distortion products. When a compressor is changing its output gain, it is acting like an AM modulator. Hence the period over which this gain change occurs will add sidebands above and below a strong signal. If this gain change occurs too quickly the sidebands will be broad spectrum and extend out past the region being masked by the ear's response to the transient tone burst. In addition to increased aliasing distortion products, if the gain change is too abrupt, it will produce a clicking sound when the transient goes through the compressor.
Hence, we need to allow the gain changes to take place over 5-20 ms in order to keep the sideband distortion products underneath the ear's masking region for that tone burst transient. That limits the sidebands to 50-200 Hz which is narrow enough to be masked by the psychoacoustic masking response of the ear.
But allowing the gain to change that slowly means that indeed we will suffer overshoots when the signal is low enough in the steady state to have some compressive gain applied. Hence the need to allow for some overshoot headroom - in this case 10 dB and more, depending on how soft the steady state is when the transient arrives.
The final control on a compressor is the release period, but this has more to do with allowing the gain to grow back to higher levels than anything else. It needs to be adjusted to complement the program content of the music. Too short and the music pumps vigorously. A little longer and it breathes. Just right and it becomes the application of mostly transparent compression.
Now on a final note... having just one broadband compressor, even if it is the composite compression curve described above, will lead to one spectral region dominating the effects at all the other spectral regions.. e.g., bass will pump the treble. This will also lead to noise modulation at those higher frequencies. The signal to noise ratio in the treble region will vassilate with the pounding of the bass, and those treble regions are far enough away from the bass that these effects will not be psychoacoustically masked by the bass sound.**
Hence what we really need are multiband compressors that all act like 2-breakpoint compressors. We need to separate the effects of the bass region from imposing themselves on the midrange and treble ranges. In this way we diminish noise modulation, the slower attacks manifest themselves with sidebands that are individually masked in each range, and the overshoot problem is contained to low level signals where it doesn't matter as much. None of the compressors will produce clipping overload on loud signal transients because those loud signals are not being compressed anyway.
... Food for thought...
**[Noise modulation -- what's so bad about this? Even in situations with low signal to noise ratios, like a faint treble with a lot of hiss background... Our brains are wired to become innured to constant irritation. After a while we don't even notice it. But what we really do notice is changes in the background. This is also true of signal phase. We can't hear absolute signal phase. But when it is changing over time we hear the familiar phaser effect. So noise modulation in the treble range due to pounding bass will manifest itself as an objectionable distraction from the treble content.]
--
DavidMcClain - 16 Nov 2004
... and here's an even better way of achieving the results discussed above for the composite compression curve. Rather than having an upper breakpoint of -10 dBFS above which no compression occurs, it is more psychoacoustically correct to have a gradual diminishing of compression at strong signal levels.
One way of achieving this is with a single Kyma compressor, put into hard limiting mode, and then blending its output with the original signal. Suppose we set the limiter threshold at -20 dB. Then the output of that limiter cannot affect strong input signals very much because it only has an amplitude around 1/10th of the input signal. Hence very little compression occurs. Down at -20 dB we have the compressor output equal in magnitude to the input signal, plus whatever compressor post gain is applied. So here and below the limiter is the main action and the input signal has less and less to contribute.
The resulting compression curve looks more and more like those old vintage compressors...
--
DavidMcClain - 16 Nov 2004
Hi David
I thought most modern compressors have delay lines in the signal path (but not in the side channel) so that the gain element gets turned down before the delayed tranisent reaches its input. Hence no big spike or overshoot at all.
--
PeteJohnston - 16 Nov 2004
