Spectral X-Contamination: Problems in Op-Amp Chips

This photograph of a classic 1960s HP scope (capable of a sharper and brighter trace than today's equipment) at Ben Duncan Research Labs shows a sinewave emerging from an innovative class of solid-state analog amplifiers, developed with Jerry Mead (founder of much-loved Rauch and Otis amplifier brands, also a record producer and drummer). The residual harmonic is so low—around 0.0001% even while delivering several amperes and tens of volts—that it contains stochastic noise, and fancy equipment isn't needed to see that Voila! we have just the 2nd harmonic. No 3rd, no 4th even. The electronics isn't single-ended and (if you read the article) clearly not complementary, and there are no transformers or tubes. "Other topologies are available." (Photo: Ben Duncan)

In high-end circles, the sonic repute of integrated-circuit op-amps (from "operational amplifier") is, at best, checkered. Of course, the expertise with which they are used and the parts they're used with can make all the difference. For example, my DIY preamplifier design, "AMP-02," published in Hi-Fi News & Record Review in 1989–90, and my earlier (1983–84) AMP-01 (footnote 1), I used the better IC op-amps of the time throughout. Both units were thought to outperform cost-no-object commercial units of the time that employed discrete transistors and even tubes, and only indicate what's possible.

What's an Op-Amp?
Nearly all the integrated circuits used in audio signal paths are general-purpose parts called "operational amplifiers." (See "Integrated Circuits" sidebar.) "Operational" doesn't mean that such amplifiers are not defective, but that, from a mathematical point of view, they're able to perform generic math operations. Amplifying is seen as multiplying; mixing (as in studio consoles) is adding; a balanced input uses subtraction; low-pass filtering is equivalent to integration; etc.

It's important to realize that not all op-amps are ICs, nor must they use transistors. The first "op-amps" were discrete circuits using tubes, and appeared during WWII in order to facilitate Project Tube Alloys—the British contribution to the Allied project to create atomic an weapon before Nazi Germany did. The US wartime invention of the op-amp by John Raggazini built on the "long-tailed pair," a powerful circuit patented in England by Alan Blumlein of EMI for the BBC's world-first TV service of 1936. After the war, George Philbrick made the first commercial op-amps with discrete transistors (see "References" sidebar [1,2]).


Whether IC or discrete, an ideal op-amp would have infinite gain, meaning that it would make an input signal infinitely bigger. Actual op-amps have hugely high gains (from 50,000 × up to 100 million ×)—way above what is of any direct use to audio, and solely over a narrow frequency range; say, up to just 1Hz, or at best 1kHz. Using negative or "regulative" feedback, this seemingly useless narrowband performance can be happily traded against a far wider bandwidth. This bandwidth can be way greater than the audioband, yet still in theory be accompanied both by any practical gain desired, usually between +6dB (×2) up to +30dB (×33), as well as by far greater linearity (lowered distortion and other errors). But not all errors are perfectly canceled.

Op-Amps in Audio
Despite their theoretical appeal, op-amps are easily and widely misused. Some early ones were not fit for serious audio applications, even when used carefully. And the misapplication of op-amps can result in unpleasant or even music-destroying sound. As a result, many audiophiles have been led to believe that IC op-amps are anathema to audio, period.

But even if your replay system contains no ICs, most professional recording since about 1972 has, for better or worse, involved passing the signal through many op-amps long before it reaches your system. Even in today's digital studios, enough of the signal path remains in the analog domain that op-amps are still involved.

A sad thing to ponder is that when you listen to many 1970s recordings (mainly rock music) on one of today's high-resolution, solid-state audio systems, the use of substandard ICs in the recording chain—or the lazy use of less bad ones—is plainly audible, once you know what you're listening for. This has left the world with a legacy of smeared, mangled recordings that, now that more than half of the 20th century's significant musical artists are dead, may never be correctable. In fact, in the four years since this article was first drafted, many musical heroes of my own—including Ian Dury, Tito Puente, Bernard Edwards, Junior Walker, and Johnny "Guitar" Watson—have passed on.

This is the point at which HLOs ("Hard-Line Objectivists"; see sidebar) usually jump in to say that, because every source of recorded music has been processed with op-amps, for someone to say that they can hear the sonic effects of a particular op-amp in a piece of consumer electronics is delusional. If op-amps are supposed to sound "unmusical," they argue, how come this unmusicality doesn't seem to apply when they're used earlier in the chain?

One answer is that the HLOs are thinking in a linear, simplistic "1+2+3" fashion. They fail to grasp that the results can be counterintuitive in a complex system in which iterative processing is used—such as subjecting music to similar reproduction errors again and again, as when the same part is used at several places in the audio chain. Thus, the result of passing a signal through many imperfect stages can be quite different from the sonic effects of any one of those stages. Example: Strident harmonics can develop into a more innocuous, less audible noise. Such noise can also then paradoxically enhance faint details in the mix, acting like dither in digital-land. (A related effect is well known in some branches of science as "stochastic resonance." [3])

Another counterintuitive truth is that measurably poor audio equipment can actually enhance music, should it fortuitously subtract from or nullify a preceding distortion. Thus, two audio wrongs can sometimes make a right. But the safer bet seems to be to begin with a near-perfect path.

A nice feature of op-amp ICs—when socketed rather than soldered directly to the printed circuit board—is that, like tubes, you can swap types or upgrade them more easily than you can discrete transistors (although, as with tubes, caveats apply). The 24 years of work I've done with various professional colleagues with recording rooms, stage sound systems, and domestic setups have convinced me that skilled and musically adept listeners readily hear the upgrading of one or more IC op-amps.

What's also interesting is that, with the modern parts released in the past decade, what's being heard is only just measurable with some of the world's most powerful audio analysis equipment. If you've read anything in the last decade about the leading edges of scientific research into human consciousness, and the holographic, higher-dimensional nature of the reality that mystics, physicists, and brain physicians seem to agree that we "really" live within [4,5], then you might find this disparity easier to comprehend.

Whether many or any of the newer, enhanced ICs (footnote 2) process op-amps appear in today's recording chains is another matter, considering how the pro-audio industry appears to be driven almost solely by the bottom line. This means that old designs that were good in their day continue to dominate the pro-audio field almost a quarter century later. (Texas Instruments' TL071 appeared ca 1976, Philips' NE5534 about 1978.) Their makers have long amortized the design costs and learned the recipe for good yields, so the low prices keep the volume up, which in turn keeps prices low. Better modern ICs have not yet managed to break strongly into this "cartel loop."

Some IC op-amps are multichannel parts, some are duals (in the same way that a 12AX7 vacuum tube is a dual triode), and some are quads (four in a package). Often, you can buy the same specification in single and dual, or single, dual, and quad packs. Dual and quad parts save some space and cost less per channel. It follows that manufacturers of audio products who listen through their balance sheets will tend to use dual and/or quad ICs in their products. And as success in the recording industry appears to be measured by how long the signal path is in your mixing console, dual and quad ICs are essential to cramming enough circuitry into a human-operable space.

Spectral & Other Testing
Since 1990 I have published a number of group reviews [6,7,8,9,10,11] on the measurable performance of IC op-amps for audio, the most recent being an update for UK pro-sound magazine Studio Sound. My op-amp testing had begun with linearity. It would have been easiest to measure the percentage of THD (vs level or frequency) as a means of categorizing sonic quality. However, this is potentially meaningless or even highly misleading [12,13]. Instead, I resolved to concentrate on showing each part's harmonic spectrum. In my experience, this information, however low in level, is far more likely to corroborate with the device-under-test's sonic character, as assessed by others before measurement.

Given manufacturers' propensity to use multichannel parts, I tested the intrinsic crosstalk (footnote 3) of a variety of dual and quad op-amps. (This is perhaps the first time this particular measurement has appeared in any publication.) I first performed a conventional measurement. A single IC test fixture previously established for spectral testing [7] was converted into a dual or quad socket on 1!9/25mm vertical extender leads. Fig.1 shows the circuit diagram; the tested IC's channel 1 (ch.1) operated as a one–op-amp balanced input stage (a differential to single-ended converter) with 0dB, or unity gain/


Fig.1 The spectral test setup, a single IC test bed modified to test crosstalk in dual and quad ICs.

The tests were carried out with two output load conditions (including the feedback resistor) of 4.3k and 600 ohms. The latter is tougher, simulating higher gain and/or heavier output loading. (The switchable 600 ohm load was "built out" with a 124 ohm series resistor so that the overall load, including the 4k3 feedback resistor, is 600 ohms.) Channel 2 (for duals; ch.3 for quads) had its output strapped to its inverting input to give unity gain, with its non-inverting input cleanly connected to ground. An Audio Precision System One Dual Domain analyzer was then connected to the ch.2 output, with the wiring heading diametrically away from ch.1, for best isolation. A ch.1 test input level of 1V RMS was chosen because it provides the lowest-distortion condition for the AP System One.

Manufacturers have long recognized the need for isolation between channels in dual and quad op-amps. A weak point is the shared power connection. Here, the ideally (but rarely wholly) rock-stable supply voltages can be polluted by the signal's demands in one channel. This variation in the supply then bleeds into the other or another op-amp channel. This effect is counteracted by a powerful feature of any op-amp, called "supply-noise rejection," which varies with different models. It's also counteracted by having a highly "active" (low-impedance) power source that instantaneously resists any voltage fluctuations. This approach was used in these channel separation tests, with a highly regulated supply situated within 2"/50mm of the tested IC's pins. This way, I could be sure that the leakage spectra seen were internal, not just an artifact of poor application.

As a precaution, I measured the residual and lab environmental noise present on the undriven channel's output when ch.1 of the dual or quad IC was also undriven (fig.2). The noisier trace of the two, with prominent spikes at 50Hz and 32kHz, occurred when the IC wasn't powered. When it was powered (lower trace), the 50Hz AC line pickup was about –135dB referred to the fundamental, while pickup of the AC harmonics—up to the 20th at 1kHz—was even lower. Therefore, the AC line harmonics would contribute nothing to the test result.


Fig.2 Residual and lab environmental noise on the undriven channel.

As another precautionary discipline, the AP test set's own residue was logged before, during, and after the testing, with the analyzer input shorted (fig.3). Note that the noise lies almost entirely 150dB below the 1V RMS test signal, excepting the apparent leakage of the second harmonic at 100Hz. This is possibly 100Hz ripple from the AP's own power supply.


Fig.3 Audio Precision System One: residue before, during, and after testing, with the analyzer input shorted.

Conventional 1/3-octave swept "crosstalk" plots (fig.4) showed that some dual op-amps have a capacitive leakage problem. If the part of the residue that's below –120dB (at 1kHz; and pro-rata below –107dB by 20kHz) is assumed to be purely random noise, then only the behavior of the TI TL072 and TL052 will be of much concern (top two traces). In these two types, the interchannel isolation is degraded to as little as –96dB at 20kHz. The sonic upset that this leads to will be greatest at HF, then in equipment with some number of dual op-amps, and will stand out as added "tizz" where the isolation between stages and channels is otherwise good. It will also be a problem when one channel handles signals that are quite different—or at quite different levels—from the other.


Fig.4 The crosstalk of 10 of the dual ICs, 1/3-octave spectral analysis (1/3-octave, non–DSP-aided).

Other ICs had considerably smaller degrees of leakage. The Harris HA5222 had the lowest, but the ranking of those at the lower levels has as much to do with having low noise as with real crosstalk. That's because the noise obscures the view—even in the analyzer's relatively selective, hence noise-free, 1/3-octave–wide bandwidth.

To better see what's really there, a tool that could eke out spectra below the noise floor is required. The solution used here is the same AP test set, working in DSP mode, with its narrower 3Hz bandwidth (a notional 1/11 of an octave at 50Hz), followed by DSP averaging—which gives stochastic-noise–subtraction capabilities (footnote 4). Sixteen samples were averaged by the DSP, which subtracts true, random ("stochastic") noise, and gives a high (better than ±1dB, or 12%) certainty to results as low as 1ppm (one part per million).

I have called this the "Spectral X-Contamination" test. In the configuration used, the AP test set permitted visual resolution of harmonics down to below –150dB, or one part per 30 million. Allowing a minimum 5dB margin above this residue to allow for the errors of interaction and ±1dB of uncertainty, crosstalk figures above –145dB are real and quite reliable.

Harmonic Issues
Before I examine the results of the test, I should briefly explain the general significance of harmonics [14]. The second, fourth, and eighth harmonics are innocuous in one sense, as their octave or multi-octave relationships with the fundamental mean that they are 100% consonant. But their presence, with possibly random phase relations to the fundamental, nonetheless retains the capacity to change the sound's timbre and pitch in ways that may be unpredictable. The third, sixth, and particularly the fifth harmonics might be pleasing in some combinations, relatively innocuous in others, and dissonant in still others. The remaining harmonics (the seventh, ninth, and all above) are almost always highly dissonant, even in tiny quantities—unless you're attuned to Japanese music.

The capacity of harmonics to compound into a morass of intermodulation products is insufficiently regarded by audio engineers. The test stimulus is just one tone, while music can contain hundreds of tones at any instant. Where you see several spikes on each graph, bear in mind that these are for each discrete frequency; at any instant, music will contain dozens if not hundreds of these, as over any period even a single instrument's simplest sound contains multiple harmonics, and hence discrete frequencies, of its own.

It's also helpful to recognize that the different tests for harmonic and intermodulation distortion are all attempts to measure just one thing: nonlinearity. The same nonlinearity that fortuitously makes only the second harmonic dominant can just as easily create other spuriae at non-harmonically related frequencies, if the second harmonic created then interacts with the fundamental or any other harmonics to produce intermodulation products. Thus, a dominant second harmonic in one stage can add third harmonic to the next, as fundamental plus second begets some third harmonic, and so on ad nauseam. In this way, playing music and hearing background grunge is one way a listener can reasonably infer the existence of harmonics that lie below the resolution of even the most powerful test gear!

Footnote 1: "AMP" is an acronym for "Analog Modular Preamplifier"—in part, a playful parody of the English QUAD name, which originally stood for "Quality Unit Amplifier Domestic"!

Footnote 2: It is little realized that IC makers are themselves dependent on fundamental research organizations. Around 1988, Bell Labs licensed its new CB (complementary bipolar) technology, which enabled vastly better-matched, better-performing PNP (bipolar) transistors to be etched into IC circuits. Before, IC designers had been hamstrung by having only decent NPN transistors, whereas discrete circuit designers had been able to use well-matched NPN and PNP transistors since the mid-1960s.

Footnote 3: As opposed to crosstalk caused by the physical (mis)layout of the circuit op-amps are employed in.

Footnote 4: Because the desired signal is self-correlated and the noise is not, summing two measurements increases the wanted signal by 6dB but the noise by only 3dB. Thus, each time the number of averages is doubled, the noise effectively drops by 3dB. This is why, when I perform a MLSSA measurement of its on-axis response to produce a loudspeaker's cumulative spectral-decay plot, I average 128 measurements, which lowers the level of the environmental noise by 21dB.—John Atkinson

CG's picture

Very interesting reading.

There's far too much to comment upon here. Not the time or place... (Not to correct, but to ask for expansion. More of this would be great!)

One comment I will make. As Mr. Duncan pointed out, IM products tend to rise and fall over real time as the relative phases of the contributing signal components change. Other effects, such as thermal changes, also create different distortion patterns over time. This is why averaging the spectrum can be misleading.

If you look at the distortion spectrum over time with a more or less real time spectrum analyzer, you can see these peaks changing. The distortion products when averaged look like bumps in the noise floor. But, if you zoom in, they're more like fields of grass swaying in the wind. This isn't unique to audio systems - you can see the same thing with various RF systems as well.

What you really hear in an audio system are the peaks of each distortion product as they change over time, not so much the averaged signal power. Listening to plain old noise is a constant change of spectral content and levels. You can hear that.

One immediate burning question - why do these design guys disappear into the aether so often?

John Atkinson's picture
CG wrote:
One immediate burning question - why do these design guys disappear into the aether so often?

From an email I received from Ben Duncan this afternoon: "I am gathering quite a list of Electronic Engineers who have been broken by the mental ordeal of designing and putting perfectionist products into production."

John Atkinson
Technical Editor, Stereophile

CG's picture

And, I'm not even in the audio biz.

But, the glory, heart felt appreciation of the customers, and the money makes up for it. Oh - wait...

Bogolu Haranath's picture

.... and all those 'electronic engineers who have been broken by the mental ordeal', are now listening to and designing, single ended, class-A, zero feedback, triode tube amps :-) .......

tonykaz's picture

I made an attempt at selling perfectionist audio products.

The perfectionist aspect is the consumer, not the Product.

Audiophiles are neurotic & psychotic.

Perfection is a moving target that Guru types review and approve ( this month ).

The headphone world seems to have an accessible and stable True-North relating to Sound Quality ( perhaps because there isn't a $50,000 Headphone Amplifier just being released )

Tony in Venice

Bill Whitlock's picture

The late Deane Jensen co-authored an AES paper (see https://www.aes.org/e-lib/browse.cfm?elib=4719) on this very subject and coined the term "spectral contamination" back in 1988. Available arbitrary waveform generators of the day had limited resolution, which limited the sensitivity of the relatively simple test. I'm hoping to revisit the subject and take advantage of the newest generation of test gear. I'm convinced that such a test is the missing link between what we hear and what we can measure because it reveals the non-harmonically-related distortion spectrum that's so irritating to our ears. For me, it also explains the widespread love of vacuum-tube and transformer containing audio gear! - Bill Whitlock, AES Life Fellow and owner/chief engineer of Jensen Transformers 1989-2014

CG's picture

First, thank you for the reply!

I *think* the problem isn't so much the generator of such signals as it is the spectrum analyzers used to measure the results. Or, at least, how those spectrum analyzers are used.

When you average a zillion sweeps to minimize the noise floor you get what we all see. Averaging reduces that noise because noise is random. But, essentially most of the potential problems then get averaged out, too. Even a 1 dB change in the measured floor because of a bunch of pseudo-random events just gets ignored. Plus, this noise floor is usually very much below other basic noise sources found in a typical listening room, or even out in the desert. So, how relevant is this averaged noise floor then, really? Humans listen to discrete events, not the average.

Now, if a measurer used a "Peak Hold" or "Max Hold" display function, I suspect there'd be more detection of audible problems. That certainly is true in communication systems. (I bet the main effect of investigating this sort of testing would be to create yet another path for decades of arguments between audiophiles...)

Of course, a suitable generator would make life as a measurer easier.

BTW, Brockbank and Wass had some thoughts on the subject back in 1945. I guess some ideas just never catch on.

Ortofan's picture

... for the mill:





John Atkinson's picture
Ortofan wrote:

From that article: "At gains of 5X and higher nothing could beat the $0.65 NE5532 except by a few dB in noise performance."

The NE5532 is the dual version of the NE5534, an op-amp that, as described by Ben Duncan on this article's "The Pink Floyd Connection" page, started life as the high-audio-quality, metal-can Mullard TDA1034. I first encountered the NE5534 in the phono stage of the Meridian 101 preamplifier and used it in a microphone preamplifier circuit I designed in the early 1980s.

Also from that article: "The older TL072, despite being used in thousands of audio devices, is way out of its league against newer op amps."

The Quad 34 preamp (or was it the 44) used TL072s. I asked Peter Walker why, when even then there were better-performing op-amp chips available? He felt that saving his customers money outweighed any measured advantage of higher-performance chips.

John Atkinson
Technical Editor, Stereophile

CG's picture
What most people measure is but a subset of the actual operating performance of devices. That's true whether it's an opamp or an amplifier in a snazzy box. There are practical limits of what a for-profit company can and will do. Really dedicated amateurs might measure more if they had the resources, which they almost always do not.

In addition, those for-profit companies have budgets for every bill-of-material used in each product. They all cut costs where they think they can, unless they can find a good reason not to. Usually, that comes down to the marketing of a product and what the market is willing to pay.

That's not a knock on these companies. For every potential customer who might say, "That's great! Exactly what I want! The price is OK, too. Sold!", there's probably 873 who will consider the price first and foremost. (That could go either direction - Inexpensive might be a negative in many cases.) I don't envy the decisions and trade-offs they have to make.

What Ben Duncan's article highlights to me is that in general, we really don't fully understand what makes a sound system sound good. There are products with relatively high distortion compared to others that are beloved based on listening. Others that would seem to be great based on the traditional measurements, but don't sound as good. If you dig deeply enough, you can usually discover why one product sounds more appealing than another. It might not be obvious, and the conclusion will almost certainly be contentious. But, you can often find it.

So, who is going to do that? They don't give out many grants for hifi performance research these days. Magazine reviewers and their editors? I don't think they have the resources. One resource being time. I also don't envy the reviewers or editors for the choices they need to make...

Ortofan's picture

... to sing the praises of the 5532/5534. His 4-part article in EE Times, entitled “Op amps in small signal audio design”, includes a survey of various devices.
He does suggest, though, that the LM4562 might finally be an improved replacement for the 5532.
Likewise, the OPA2134 is suggested as an upgrade for the TL072.

The QUAD 44 preamp used many TL071 op amps – the single version of the TL072.

One advantage of the single 5534 is that the compensation pins are brought out, which enables the possibility of bypassing the input stage for applications where lower input bias current is needed.
An old Siliconix FET databook includes an application note illustrating such a configuration. The normal input pins are connected to the negative supply rail and the bipolar input stage is replaced with a cascode connected pair of 2N5912 dual FETs. The balance and compensation pins then become inputs to the second stage differential transistor pair.
While the 2N5912 no longer seems to be readily available, the Linear Systems LS5911 is a substitute.

One wonders if it ever occurred to Peter Walker to offer a “deluxe” version of his preamp with better performance, from the use of higher performance devices, at a commensurately higher price?
He might have responded that the existing design already performed as a “straight wire with gain” and asked what more than that did you need?

invaderzim's picture

"This has left the world with a legacy of smeared, mangled recordings"

It is so sad that many of us will never know what these performers really sounded like. The true performers from when our equipment just wasn't up to the task of capturing their sound.
Paintings can be restored, manuscripts can be repaired and duplicated but we can't clean or recreate the sound and have it be correct.

CG's picture

Perhaps - some day.

A bigger concern might be the care, or lack thereof, the keepers of the master tapes have shown over the decades. A lot has been thrown out and a lot has been lost to fire. There's some indications that the recording companies have invested a bit more to keeping the recordings - the ones they make their money from, which is crazy if you think about it - stored better. But, an incredible amount has been plain lost forever. I'm sure that has been the case with written documents and other forms of art, but the record companies haven't even been willing to protect their own assets.

CG's picture

I wrote a considered reply to a comment above. Nothing remotely derogatory, inflammatory, or impolite. No bad words.

Gone! It was saved, as shown by the blank text box.

Oh well. This may be a message from the cosmos that posting comments on the internet is not for me.

John Atkinson's picture
CG wrote:
I wrote a considered reply to a comment above. Nothing remotely derogatory, inflammatory, or impolite. No bad words. Gone! It was saved, as shown by the blank text box.

I don't know why the text disappeared but I inserted it from the email I was sent alerting me that there had been a response to one of my postings.

John Atkinson
Technical Editor, Stereophile

CG's picture


I still think the cosmos is trying to tell me something though...

tonykaz's picture

it happens, from time to time.

You can go back a page, in your computer, to retrieve it and re-submit.

It only happens in Stereophile Comments.

Tony in Venice

sdunbar's picture

As an Applications Engineer for Texas Instruments, I can tell you that with advancements in process (smaller die) and packaging (smaller packaged devices) there is very little reason to use dual or quad packaged op-amps in pro-audio applications.

Yes, one can still save some money using dual/quad packages, but the need to use these packages to save space is relaxed.

The OPA1641 that is now already 10 years old and available in a small VSSOP package could be a good start. Other very capable devices are in smaller packages, such as SOT-23. There are literally 100's of devices to choose from, and there is no need to use exactly the same op-amp throughout an entire signal chain. Part of the discretion of design is to use the right part in the right place. Some areas maybe you can cut corners if you know what you're doing, others you cannot.

Linearity is of course the key to good reproduction. I have yet to see a widely accepted definition of a test for characterizing this parameter. Something based on a comb of multiple, yet non-harmonically related tones and then measuring the residual harmonic and IM distortion and expressed as an RMS value might be a good start. It would be necessary to reduce the tone amplitudes so that the peak/average ratio of the combination is still within the boundaries of the op-amp. But clearly there is plenty to argue about here.

CG's picture

There is this:


A test like this lets you examine the odd and even order IMD products. The fundamentals aren't eliminated, though. That probably could be handled through software trickery.

I'd think that an even more complete test would be to weight the levels of the applied test tones spectrally, so as to simulate some facsimile of the music/sound track/spoken word power distribution. There's higher levels in the lower frequencies than at 20 KHz. Why not shape the test signal? At the analysis end, you could also potentially use A-weighting to approximate the human's ear's sensitivity.

In addition, using some kind of peak hold measurement so that the maximum value in each FFT bin is saved would be desirable. Although averaging helps you look at very low levels because of the randomness of true noise, it also misses distortion peaks that occur when the relative phases of the distortion tones line up to add. You'd hear these peaks, but an averaging scheme would, well, average them them out to be less representative of the reproduced sound.

As for whether it would be widely accepted, I'd doubt it. You can't get anybody to agree on anything. Don't you agree? :8^)

sdunbar's picture

If one knows where the fundamentals are (because the test gear is generating them) and there are no FFT leakage artifacts, it should be simple matter to arithmetically subtract out the fundamentals, leaving the HD and IMD products behind.

Notice how far they have backed-off the magnitude of each tone? No doubt to manage the crest factor/peak-to-average ratio. I would think there are diminishing returns to adding more and more tones; you can argue that more tones better represents real audio, but at the same time you still need to be able to accurately extract the fundamentals and resolve the FFT.

Real audio tends to have lower amplitude content at higher frequencies so using the same amplitude for every tone is probably not the right thing to do, either.

All of this needs to boil down to a single "number" that can be used for comparison purposes. So, RMS of residual.

Surprised that IEEE hasn't standardized something like this decades ago. In a sense, audio is such a commodity now it really SHOULD have industry standards. Since audio is so dependent on linearity, seems like an obvious parameter to quantify.

But your right... it's hopeless.

CG's picture

In the CATV world they've long relied on multi-tone testing to determine the linearity of their systems. That analog testing has been done for at least 40 years. There is CTB - Composite Triple Beat - for odd order IMD products and CSO - Composite Second Order - for even order products. Plus SNR. Three numbers, but people understand what they mean. Today, that's gone away a bit, since the digital TV channels are QAM modulation, so the distortion artifacts are a little different than they were for plain ole analog TV channels. But, the new tests are similar.

In ye olde analog telephone days, the multiple analog channels were frequency division multiplexed so that you could get more than one phone call per line. In fact, they used groups that would let them get hundreds. There, the testing was done with white noise, with some small frequency bands notched out. During testing the noise power in the notched-out bands was measured and compared to the overall power, giving a single performance number called NPR - Noise Power Ratio. IMD of white noise is more white noise, so this gave them a sophisticated way to measure not only the distortion of a group of channels, but also the basic noise level. They pretty much stopped using analog phone channels in the 80's sometime. They were performance testing way before that.

I don't know what the answer is. Perhaps somebody ought to just put a stake in the ground for some sort of test and see how it progresses. The problem then becomes, just what can we hear and what is bad for sound reproduction?

sdunbar's picture

It’s impossible to argue with someone else about THEIR hearing ability. Certainly doesn’t work with my elderly mother-in-law living with us.

I'm the first to admit that my own ears are probably not that attuned. While my own mother had a degree in music and was a piano teacher, I completely squandered that opportunity... nevertheless, I did listen to quite a bit of audio growing up.

Not to be an "HLO", but my general assumption is that if someone/anyone can hear some difference, then that difference can be measured since we can build test equipment that is far more sensitive/perfect than our ears. This article seems to support that assertion.

What’s aggravating is when such a claim is made, but then we’re just expected to take the listener’s word for it- show me on an analyzer, please.

This is further obfuscated when certain distortion products are considered "pleasing" to some people. It's fine if you like the sound of second harmonic distortion, for instance, but from a recording and reproduction perspective, it's still distortion.

CG's picture

Two things...

One is that being even a virtuoso on an instrument, or many, is no guaranty that you're a good listener. My wife went to a fancy music school, played with and was tutored by several recognizable musicians there, and so on and so on. She tells me that most serious musicians aren't concerned with fidelity when they are listening. They've certainly listened to a lot of lousy sounding music in bad halls, just as they've listened to great sound. For them, they hear the musical intent in their head. They fill in the blanks. So, don't feel badly about that part.

(Fun fact: After getting critical, umm, "compliments" from their teachers, most of the voice majors would go sing to themselves in the porcelain lined rest rooms, just to make themselves feel better. Most everybody sounded glorious in there due to the natural reverb. Not sure if the tuba players did that. The pianists certainly didn't.)

Second thing is that just because it's probably possible that we could measure whatever distortion if we only tried, that's not what's done. Test equipment is good in that it often measures one characteristic at a time. But, to do that, it often trades off sensitivity for some of the bigger picture. A good example is the use of averaging and very narrow FFT bins to be able to capture teeny tiny harmonic products. As you pointed out, that not only doesn't exercise the DUT with regard to crest factor considerations, but the distortion products that are subject to the moving peaks are deliberately ignored. Can your hearing system do that?

sdunbar's picture

Certainly this has been discussed ad nausium elsewhere, but whether or not something "sounds good" or "sounds xxx" where xxx is your favorite audio related adjective, will always be a subjective matter.

For it to be objective, it would have to be time and space invariant, and since we cannot transplant ears, etc... from one person to another, it will remain subjective. Even the most attuned, demonstrably astute, refined listener is still engaging in an experience that by its very nature is subjective, and their descriptions of the experience must then be so as well.

Everyone also has their own tastes that can be tempered only so much, and it just becomes more subjective.

There is nothing wrong with this; that's just how it is. There are things about an audio signal chain that may be objective. Linearity performance, cross channel contamination (hence the article), noise floor, power consumption, etc... all of these things can be objectively compared through experiments that can be reproduced at will.

While nothing can be "perfect" we can certainly compare two specimens to see which is better in any particular dimension. My fear is, and this is probably why so many of these audio electronics engineers have vanished, is that despite desirable and measurable improvements in signal chains, there's still some "audiophile" who just doesn't like how it sounds.

When it become irrational, it's no longer fun for the engineer.

CG's picture

I think that when it gets in the way of making a living, that detracts from the fun in a big way.

Audio is not my livelihood, so that's easy for me to say.

All that said, I think there's still more that could be measured in a way that would help give better definition of what is better. Obviously, we're not there yet.

For me, it's far easier to make adjustments and then measure through listening. Jumping to the end is probably not right, but it's hobby.

Herb Reichert's picture

And bravo (!) Ben Duncan for such a broad, knowledgeable, cultured view of some very complex (and thorny) topics.

I am envious of your skills.


Ortofan's picture

... "Hard-Line Subjectivists"?
By not doing so, is he revealing his own biases?

Worth reading is the referenced letter from Bob Orban (split over two pages):

John Atkinson's picture
Ortofan wrote:
Worth reading is the referenced letter from Bob Orban (split over two pages): https://www.stereophile.com/content/scientists-ivsi-audiophiles-1999-more-letters . . .

Also in the referenced article, a superbly written letter from the late Harvey Rosenberg: www.stereophile.com/content/scientists-ivsi-audiophiles-1999-more-letters-part-4.

John Atkinson
Technical Editor, Stereophile

CG's picture

Dr. Gizmo sure was great!

I'm not sure that digital can't sound great, but an awful lot of recordings that were digitized were done in a hurry, with not much care as to the sound. Or, maybe worse - some, ahh, creative engineer (probably coked up to keep going at 3AM in order to get the releases out quickly) decided to add his or her own special interpretation of the original artists' and producers' work.

Ortofan's picture

... the "Digital Pact" neglected to ensure that the agreement included the recipe for second harmonic sauce.

Maybe Dick Burwen and Bob Carver should collaborate on a device to add transient and background noise into digital recordings to make them sound more like analog.