A Discussion About True Power Summing for Stereo Compressors

Where we discuss new analog design ideas for Pro Audio and modern spins on vintage ones.
Post Reply
User avatar
mediatechnology
Posts: 5437
Joined: Sat Aug 11, 2007 2:34 pm
Location: Oak Cliff, Texas
Contact:

A Discussion About True Power Summing for Stereo Compressors

Post by mediatechnology »

This was originally published by me in the Prodigy-Pro forum.

The Pico Compressor Stereo Edition will be using a detection method called "True Power Summing." I thought it would be a good idea to re-visit the subject as it is one of the key design elements of the Stereo Pico Compressor and is a build option for Original Pico Compressors in linked-dual-mono.

The Blackmer (tm) RMS detector, used in the THAT4301, THAT2252, and most dBx products provides a measure of signal power. When a stereo compressor using individual rms detectors in the left and right channels have their rms outputs combined at the timing capacitor, the powers sum. Remember that a doubling of power is a 3 dB increase.

In an experiment I ran recently I combined the detector outputs of the Pico Compressor. Each, with 1V (rms) audio in, produced 1VDC out (after trimming them to match). Combined their outputs were ~1.019 VDC. The output scaling of the THAT4301 is ~6.6 mV/dB. Thus the combined detector output produces an ~19.8 mV increase representing a "true" 3 dB power sum.

We discuss this again in Keith's Ultimate GSSL post. In Keith's Post we learn that his Ultimate GSSL is more similar to the orignal SSL Mix Compressor in that it detects whichever channel has "the greater of two" peaks. (And I might add that his differential VCA connection is brilliant.)

Edit 11/07: I also wrote this article published here in the Pico Compressor Forum:
http://www.ka-electronics.com/forum/php ... p?f=6&t=17

The original GSSL Clone posted here in this forum uses an algebraic sum where the left and right inputs are combined as audio prior to detection. This can also be called L+R detection. It is an option available to the Pico Compressor builder as well. L+R detection is computationally biased to detecting, and thus more heavily compressing, mono. If two identical inputs are fed into an L+R detector, but one of the inputs has its' polarity reversed, the detectors output will be zero because the sum of the audio input is zero. Thus, more highly correlated signals will receive the most compression, while completely uncorrelated material will receive no compression. Put simply, mixes get wider.

Keith's Ultimate GSSL and the Pico Compressor combine detector outputs. Because of this, both the Ultimate GSSL and the Pico Compressor are insensitive to phase, polarity, or correlation because summation occurs after detection. In my previously cited experiment I had great fun flipping signal polarity and seeing no combined detector output change.

So what are the differences between true power summing and "greater of two" peak detection? I'm willing to go out on a limb here and make a broad generalization: Peak response is very desireable when the signal has to fit the confines of a transmission channel or storage medium. It provides fast response. RMS detection more closely emulates perceived loudness and is thus more suitable to increase program density. Think 1176 vs. LA-3. The Stereo Edition of the Pico Compressor will have multiple "peak" and rms timings.

Both peak detection and rms have their place. "Greater of two" peak response, or true power summed rms, will both image more accurately than an L+R or algebraic sum will because they are phase and polarity insensitive.

I'd played around with discrete rms detectors many years ago and remembered that the outputs could be combined with one another. After I proposed combining the detector outputs of the 4301 for the Pico Compressor and confirming that it was possible, I discovered that it had a name: True Power Summing. See THAT's application notes on the subject DN-116 (page 4) and DN-118 (page 2).

So after I received my Pico Compressor Boards I built one:
Image
And it sounded incredibly smooth! So I decided to record a demo cut with all controls set to maximum compression to see how it would image. (Threshold Min, Ratio Max, Makeup Max. Hammered real hard.)

Listen for yourself: https://www.ka-electronics.com/Content/stereodemo.mp3

Vocal stays out front eh? Close your eyes and it paints a picture. Now it is squashed I'll admit. But things don't move. As you can see there's lot's of wires though even though the true power sum portion is nothing more than a short between timing caps. (The slave channel is on the right.) Since my entry wouldn't win any beauty contests Roger built his. Checkout those Blue LEDs.

And he built it with True Power Summing as well. Now each of us had to do a lot of extra work (OK Roger had to do a lot of extra work) to enable the switching between dual mono, stereo etc. It added a lot of time to the build. And he listened a lot. I think that's when he decided it might be good idea to build a dedicated stereo compressor that used true power summing.

So who else has done this? Well THAT knows about it and it goes as far back in dBx as the 160. But strapping a pair of 160s together is something I almost never saw. We've also seen references to it in the 165, 166 and the dBx Quantum whose rms calculation is done in DSP. We find these in various dBx manuals:
Image
Image
Image

And also this:
http://www.stellacustomelectronics.com/c1.html
A mastering compressor? Cool.

Edit 11/07: Apparently Stella Custom Electronics has ceased production.

The Pico Compressor Stereo Edition will not only have True Power Summing but, like the Original Pico Compressor will have a switchable Peak Mode along with a new adaptive Auto mode using the "non-linear capacitor" as featured in THAT's DN-114.

Stay tuned...
User avatar
JR.
Posts: 3697
Joined: Sat May 24, 2008 7:21 pm
Contact:

Re: A Discussion About True Power Summing for Stereo Compressors

Post by JR. »

I'll chime in some more tomorrow.. I have been trying to decipher DBX's claims for RMS vs. simple average for decades. They were making claims about it impacting decode accuracy with their tape NR, back when that was a major business.

I even recently did some work with RMS vs, Average detection in my Peak/VU meter..

More later...

JR
Cancel the "cancel culture", do not support mob hatred.
User avatar
JR.
Posts: 3697
Joined: Sat May 24, 2008 7:21 pm
Contact:

Re: A Discussion About True Power Summing for Stereo Compressors

Post by JR. »

Before I start ranting about RMS, lets not forget important rule #1 about audio.. If it sounds good, it is good.

OK where to start. Regarding the sound of dynamics processing, we are basically multiplying an audio waveform by a control voltage. The spectral content of this control voltage can add artifacts to the resultant audio. A certain amount of masking goes on around dynamically changing music signals so we play games with attack and release times on this control voltage to mitigate audibility of these modulations.

How does RMS factor into this? First I would like to remove peak detection from this discussion and reduce it to a comparison only between RMS and simple rectified average. I use that specific verbiage because unrectified average of an audio waveform is 0V. :lol: The basic difference in RMS is that the rectified signal is squared, or multiplied times itself, before it is integrated, or smoothed by a filter, and after the smoothing, the square root is taken to normalize level. While it seems from inspection of the math this should make a dramatic difference between different waveforms, in practice using the longer integrations times appropriate for viewing or processing i don't observe significant differences.

Data point #1. A few months ago I coded up a RMS routine inside my peak/VU LED meter, basically using true RMS for the rectified average in the VU portion of the meter. Using the exact same integration time constant, I ran two meters side by side on a range of music and could not detect a difference between them. Note: In hindsight I didn't mathematically null the two results to display only the difference, that might be interesting if/when I revisit this.

Data point #2: This goes back to the old Tape NR products. DBX based marketing claims on the superiority and uniqueness of their companding NR products because of "RMS" detection. I made NR designs based on Signetics NE570 series, but with custom and pretty sophisticated detection schemes to deal with idiosyncrasies of tape. The anecdote that relates to this discussion is not my design work, but a friend's company who chose to make and sell an outlaw DBX compatible NR based on the NE570 series. My friend invested bench time in tweaking the attack/release time constants to null his companders against DBX units, as I recall he got bench nulls down to around 40dB or more, which suggests to me that RMS differences are at best very subtle.

Next, I will discuss combining multiple audio streams for dynamic processing. I have done a certain amount of work in this area for automatic mixers. I even have one improvement patent building upon Dan Dugan's classic work in this area. Bear with me here, since I believe this may apply to the current discussion. The classic Dugan gain sharing algorithm involves comparing individual inputs to a combined sum, to compute gain distribution, so it will always sum to unity. For this particular example it is critical that the raw audio waveforms be combined before rectification and logging for gain calculations. If the signals are rectified before combination you lose "information" namely about the coherence of the inputs. To wit if the same signal is detected at two inputs with equal level the output will double. Two equal level incoherent signals will only increase by the square root of two or +3dB. I recently helped some guy who was trying to code a Dugan style automatic mixer to work as an add on to digital mixer, but he only had access to meter levels, which were rectified. His "Dugan" algorithm didn't work properly because it couldn't discriminate input signal coherence.

Now to get to my point, when we compress stereo audio paths, it is popular practice to link the two channels. In effect making both channel gain modulations identical to keep stereo imaging locked dead center. So we are in effect combining control input from both channels. The important issue IMO, since I am downplaying the significance of RMS, is whether our combing approach maintains or discards signal coherence information. I will qualify up front that compression is an effect, so there is no right or wrong result like there is in automatic mixing. That said using "RMS" detection (or IMO any rectification) before summing, will over report incoherent signals relative to coherent.

In typical panned stereo mixes the vast majority of audio is coherent, so if the signal envelope is dominated by only coherent information there will not be a huge difference, but I will predict a 3 dB difference for a hard panned full left/right signal between AC summing or DC summing.

If looking for features to perhaps include in a future comp, consider also offering the option between AC and DC combining of L/R. I predict this will offer a subtle difference between treatment of mono and stereo information. Over reporting the incoherent should increase apparent stereo effect.

Some day if i ever get around to making my uber-comp, this will be just another variant added to my long list.. but my comp is just mental masturbation that I may never finish. You guys are in the middle of making sounds, so I hope some of my comments make sense and are useful.

JR
Cancel the "cancel culture", do not support mob hatred.
User avatar
mediatechnology
Posts: 5437
Joined: Sat Aug 11, 2007 2:34 pm
Location: Oak Cliff, Texas
Contact:

Re: A Discussion About True Power Summing for Stereo Compressors

Post by mediatechnology »

That said using "RMS" detection (or IMO any rectification) before summing, will over report incoherent signals relative to coherent.
Thanks for the discussion John. I agree with most everything else except the above.

I believe the opposite of the above is true: Summing prior to rectification will over report coherent signals. Equal level signals of opposite polarity (about as incoherent as it can get) when summed prior to detection will produce an output of 0V. Keith illustrates that in his ssl turbo mod video where he shows a GSSL doing exactly that: No GR when equal inputs are applied with opposing polarity. The GSSL is well known for punching a hole in center due to coherent bass. More coherent = more GR.

If summing is done after detection, specifically true power summing, input phase angle or coherence doesn't affect the measurement. (As long as the averaging time is sufficiently long). Equal RMS level inputs increase the output measurement by a 3 dB power sum. I can send you a Pico Compressor or a pair of 2252 RMS detectors and you can hook them up as I did with a pair of generators to see for yourself. I don't think that's arguable.

My point in posting that article was for the true power sum discussion. I'm not very concerned about tape noise reduction or whether an NE570 makes a great compander. However, there are a great many 4300-series companders in wireless mics and for a part that's several times more expensive than a 570, there must be a solid reason why they're used. Cordless telephones have kept the 570-series economically viable.

What Roger had found with the A/R board (which makes it no longer RMS) is that he can use it to tailor compression and surgically alter stuff making it more of an effect. I believe his conclusion was that for "set and forget" natural-sounding applications not needing repair he just couldn't beat RMS or RMS with non-linear integration (dual time constant).

I think rule #1 applies. To most people's ears the TPS/RMS combo works very well. We've only had one person say that they didn't like the sound of the Pico. They admitted that it was bought second-hand and they weren't really sure it was assembled properly. Can anyone point me to an NE570 compressor that's in the pro market? (not MI).

Roger - Are you available for comment?
User avatar
JR.
Posts: 3697
Joined: Sat May 24, 2008 7:21 pm
Contact:

Re: A Discussion About True Power Summing for Stereo Compressors

Post by JR. »

That said using "RMS" detection (or IMO any rectification) before summing, will over report incoherent signals relative to coherent.
Thanks for the discussion John. I agree with most everything else except the above.

I believe the opposite of the above is true: Summing prior to rectification will over report coherent signals. Equal level signals of opposite polarity (about as incoherent as it can get) when summed prior to detection will produce an output of 0V. Keith illustrates that in his ssl turbo mod video where he shows a GSSL doing exactly that: No GR when equal inputs are applied with opposing polarity. The GSSL is well known for punching a hole in center due to coherent bass. More coherent = more GR.
Aren't we saying the same thing/?

---------
No need to get defensive , my comments about 570 based designed was all about detection schemes not the gain elements themselves. FWIW I designed a compressor back in the early '80s where i used a NE572 as a cheap FW rectifier in my sidechain, to control Valley Audio (Buff) VCAs.

I don't intend this as a criticism of your compressor (while I remain unsold on RMS as some secret sauce). I am just suggesting that the rectification which conceals coherence information is the more important mechanism (IMO), not squaring the waveform before applying time constants.

To revisit my comment about AC or DC summing: For two coherent signals, which is what we get from panned mono sources, they will combine similarly AC or DC. True stereo sources which are not identical, will destructively combine in AC summing, so DC summing will make them appear stronger than AC summing.

My point is that true streo information is enhanced relative to mono information by rectification prior to combining. This is just another tool in our tool kit we can use to get different results from a dynamic processor.

JR
Cancel the "cancel culture", do not support mob hatred.
User avatar
JR.
Posts: 3697
Joined: Sat May 24, 2008 7:21 pm
Contact:

Re: A Discussion About True Power Summing for Stereo Compressors

Post by JR. »

raf wrote:Almost forgot to mention this..

Back when I got a decent 4 track, a Tascam reel to reel @ 15 IPS, I built a 4 channel version of JRs compander from the 1981 Popular Electronics article and used it blissfully ignorant of higher priced NR systems. Sounded very good and I was happy for a while... Then I decided to take the plunge and buy 4 channels of dbx 150 NR for my deck... You know, get more professional...Well, I didn't like the results as much as I liked the lowly NE570 based version that JR did. I spent so much time trying to optimally set the 150s up that I ended up not doing very much recording. I sold the 150s and the Tascam and went to DAW based on the Atari Falcon (Motorola '030) and German made I/O. Eventually went to Mac based DAW with MOTU 24 bit I/O where I am today.

I still have that 4 channel compander in storage somewhere. I did a layout for a single channel of compander and hand etched them based on the schematic in the article. I remember buying several of JRs PCBs back in those days... EQ, Stereo NR, BBD delay etc. I really enjoyed those days!

rf
In the context of dynamics processing the audibility of gain modulations will typically trump differences between gain elements, but all thing equal improving any link in the chain is worthwhile. The classic 570 gain cell had its bad behaviors. Trust me I spent years trying to mitigate the changes in CV feed through with level and other irritating characteristics.

My original 1977 NR kit was really a very basic data sheet implementation of the ne570 chip. The only cleverness I added, was a dynamic HPF. I made a level adaptive one pole HPF by using a fixed value input capacitor, with the resistor connected to the inverted output of the 2:1 compressor stage. The pole frequency shifted with the gain of the compressor, so at very low level, when gain was highest, the HPF shifted up in frequency. This had the very real benefit of scraping LF "out of round" or warped record modulations out of the signal envelope at very low levels, where they could cause phantom mistracking modulations on playback, when the tape recorder didn't return the LF warp signal. At higher levels where the envelope was dominated by in band signals the filter was opened to full bandwidth.

I am more proud of my later tape NR that I did several years later in the '80s. This used the NE572, and I broke out the two rectifiers for far more than just different attack and release, but level dependent attack/release and other tweaks... This was all a waste of my time since customers could buy DBX NR fully assembled in Japan, for the price of my kits. Ouch..... IMO this later offering was better, despite the inferior gain element, but it's all moot since DBX has the brand power, and people mostly bought my kits to save money.

JR
Cancel the "cancel culture", do not support mob hatred.
User avatar
mediatechnology
Posts: 5437
Joined: Sat Aug 11, 2007 2:34 pm
Location: Oak Cliff, Texas
Contact:

Re: A Discussion About True Power Summing for Stereo Compressors

Post by mediatechnology »

Aren't we saying the same thing?
Not exactly, John but close. Summing after detection, at least with rms, washes out coherence/incoherence altogether. It doesn't emphasize or de-emphasize those components. They just don't matter.

I think I would agree to: "That said using "RMS" detection (or IMO any rectification) before summing, will report incoherent signals the same as coherent."
My point is that true streo information is enhanced relative to mono information by rectification prior to combining. This is just another tool in our tool kit we can use to get different results from a dynamic processor.
I agree. I think that job is better served by an M/S matrix, where one can chose which domain to process, as part of the sidechain and or/audio path. I think the primary interest in our M/S encoder/decoder is M/S ratio processing via dynamics. With that approach one can have it either or both ways.

My point about the 570 or any other averaging detector being made to emulate rms is that it seems to be a lot of effort just to avoid something you can now readily buy. When a dbx license was required it made a lot of sense to avoid those patents. It doesn't now. $9.95 gets you a VCA, an RMS and op amps. I'm not sure rms is magic Mojo where peak level control is required, but for density improvement it's pretty convincing to the ear. It's unusual to hear it make a gain change where it seriously f**ks up. Although one could dial in averaging time-constants to do the same thing, why bother?

Roger wrote, refering to the GSSL:
I could not get around "mono deafness" (hole in the center) while summing ahead of the rectifier.
I think it's actually the other way around. Being an L+R (pre detector sum) it only "hears" mono and thus acts upon only mono. It's "deaf" to stereo. I mention this due to the irony. The stereo bus compressor of choice for the "paint-by-numbers" crowd is actually deaf to stereo.
User avatar
JR.
Posts: 3697
Joined: Sat May 24, 2008 7:21 pm
Contact:

Re: A Discussion About True Power Summing for Stereo Compressors

Post by JR. »

mediatechnology wrote:
Aren't we saying the same thing?
Not exactly, John but close. Summing after detection, at least with rms, washes out coherence/incoherence altogether. It doesn't emphasize or de-emphasize those components. They just don't matter.

I think I would agree to: "That said using "RMS" detection (or IMO any rectification) before summing, will report incoherent signals the same as coherent."
My point is that true streo information is enhanced relative to mono information by rectification prior to combining. This is just another tool in our tool kit we can use to get different results from a dynamic processor.
I agree. I think that job is better served by an M/S matrix, where one can chose which domain to process, as part of the sidechain and or/audio path. I think the primary interest in our M/S encoder/decoder is M/S ratio processing via dynamics. With that approach one can have it either or both ways.


My point about the 570 or any other averaging detector being made to emulate rms is that it seems to be a lot of effort just to avoid something you can now readily buy. When a dbx license was required it made a lot of sense to avoid those patents. It doesn't now. $9.95 gets you a VCA, an RMS and op amps. I'm not sure rms is magic Mojo where peak level control is required, but for density improvement it's pretty convincing to the ear. It's unusual to hear it make a gain change where it seriously f**ks up. Although one could dial in averaging time-constants to do the same thing, why bother?

Roger wrote, refering to the GSSL:
I could not get around "mono deafness" (hole in the center) while summing ahead of the rectifier.
I think it's actually the other way around. Being an L+R (pre detector sum) it only "hears" mono and thus acts upon only mono. It's "deaf" to stereo. I mention this due to the irony. The stereo bus compressor of choice for the "paint-by-numbers" crowd is actually deaf to stereo.

We're getting closer on the semantics. :D

My whole point for raising the audible difference between rectification pre or post combining is looking for a plausible explanation for differences heard attributed to RMS.

I expected a dramatic, at least visible difference between RMS and average response in my meter or I wouldn't have invested the time and effort to write the software. The hardest part was writing a fast square root routine. I have rolled my own RMS conversions in the past with hardware but I always just accepted the superiority of RMS as a given, I never compared it side by side. The beauty of my comparison in the digital meter is that even the time constants are matched to a precision difficult to attain in the analog domain.

I regret now that I didn't take it to the next step and null the two data streams to see how little they diverged (they must be different right?) and what kinds of music caused how much difference.

I've got better things to do this weekend than write software, so I will leave this "how is RMS really different? discussion to be continued. My sense for now is that the long time constants relative to waveform period reduces the practical differences. If for DC (infinite time constant) average and RMS converges, the question becomes at what time constant does any difference express.

Even if I am correct (stranger things have happened) this doesn't advise against using the THAT chip set. It's a PIA to do the log conversion to get audio level in V/dB, so their chip set is probably the cheapest way to generate that data to interface with their LOG input VCA, RMS or not.

I apologize if I sound argumentative, I have been pondering claims about RMS for tape recording NR accuracy for decades. Since the RMS is not patented or patentable, I could have used it if I were convinced of a benefit. I found other more serious issues there.

I was ready to put it into my PK/VU meter until I didn't see any difference. When the result is the same use the less complex approach. I am tempted to put it in a later version as a marketing hook, sInce i've done most of the work. The "force" is strong in this RMS.... 8-)

JR
Cancel the "cancel culture", do not support mob hatred.
User avatar
mediatechnology
Posts: 5437
Joined: Sat Aug 11, 2007 2:34 pm
Location: Oak Cliff, Texas
Contact:

Re: A Discussion About True Power Summing for Stereo Compressors

Post by mediatechnology »

John - Thanks for your thoughtful posts. I think we're "converging."
My whole point for raising the audible difference between rectification pre or post combining is looking for a plausible explanation for differences heard attributed to RMS.
There are audible differences in mono compressors WRT averaging vs. RMS detection where L/R combining isn't even a factor. The mono case eliminates a whole 'nother set of variables.
I expected a dramatic, at least visible difference between RMS and average response in my meter or I wouldn't have invested the time and effort to write the software.
What musical sources did you use for the comparison? Was it commercially-recorded music that had been processed and mastered?
The hardest part was writing a fast square root routine. I have rolled my own RMS conversions in the past with hardware but I always just accepted the superiority of RMS as a given, I never compared it side by side.
RMS and average in DSP do have the advantage of being numerically precise but RMS is indeed process intensive. As a benchmark, how do the number of code lines compare?
The beauty of my comparison in the digital meter is that even the time constants are matched to a precision difficult to attain in the analog domain.
Yes, analog can do the logging quite easily but things get loose when, resistance, capacitance and temperature enter the picture. TI recently wrote a paper - I think it was geared to medical device implantables - where they compared analog vs. DSP processes. Although DSP was more precise, the conclusion of the paper was that analog should never be ruled out in implantable applications because it won on power consumption. MIPS and heat vs. analog tolerance.
I regret now that I didn't take it to the next step and null the two data streams to see how little they diverged (they must be different right?) and what kinds of music caused how much difference.
I think that's one of the keys. They are going to be different in some way though how significant depends. Under steady-state they can be the same. (The low-end avg responding DVM with RMS calibration.) Based on what you're saying and the noise reduction experiments you did, the attack and release characteristics, can be made close.

But what about crest factor error? "What kinds of music caused how much difference?" One of the things Blackmer mentions early in the patent is crest factor. We know that in instrumentation high crest factor signals produce large errors when measured with average-responding instruments. Blackmer also discusses the need for power measurement to emulate hearing.

When measuring high crest factor signals average responding meters under report power. It would seem to follow that average responding dynamics processors would also under-report power and thus under-compress with high crest factor signals. The more "peaky" the signal the less they do.

I think that with unprocessed recordings you would have seen differences in meter indication. Few modern CDs, broadcasts or even raw tracks from sound libraries and sample files escape processing and compression. Most are squashed to death and many are fully-clipped. The square waves found on a modern pop CD essentially have unity RMS, average and crest factor values. Looking at any mildly processed track I doubt you would find an element with a crest factor more than 2-4.

Hang a mic up in the studio and blow a sax and you'll see crest factors of 8-10. Voice is about the same. I'm of the opinion, (and this thread has caused me to think more about it than I have in awhile thank you John) is that the high crest factor of raw elements may be one of the primary differences in RMS vs. average detection schemes. RMS is going to more accurately measure power, and thus satisfy the ear, with raw, real-world music.
Post Reply