Ethan Winer Builds a Wire Null Tester

[solderdude]

of course he did not use a high-Z source.

@Ethan Winer wrote this about the null tester:

The Completed Tester
Once the reduced Null Tester was completed, I made a video to demonstrate the device in operation. The first part of the video explains all about null testing in detail, then the demonstration portion compares three inexpensive wires plus a $700 “boutique” wire. I don’t mind giving away the ending: All four wires were proven to pass audio identically down to a level below –110 dB. In fact, I didn’t think to make this important point in the video: The Null Tester’s inherent noise floor is –110 dB, but our ears can discern the presence of music and speech as soft as 20 dB below the noise. When the Null Tester is set for the deepest null, no music can be heard at all. That means the null is actually 130 dB down, or even more, not “merely” 110 dB. I wish I had thought to make that point where I show music being nulled out completely in the video!


I see no -84dB and no 14 bits there. It really does not matter if the source is just 10 bits or 24bits in resolution.
They will both be able to null down to -110dB (noise floor from tester)

 

[solderdude]

I said High Z device, which refers to his null tester and amp after it. His curcuit board shows he has a 1K pot for splitter box, then a 1.1K resistor for null tester input. This is going to be a load of minimum 1.05K for the sansa.

I have no idea what the input resistance is of his null tester (do not have the complete schematic) but even when it is 1k then this is absolutely no load for a Sansa which can effortlessly drive 32 Ohm. Besides distortion of the input and noise levels are of no consequence for the nulling process at all.
I would be really very surprised when Ethan would have used different input buffer circuits. It's the input buffers that load the music source.

Then for his null tester, opa1612 has a gain of 1. Channel 1 Outputting to a load of 5.1K before 80-100 dB amp, while channel 2 has another 5.1K before inverting opamp (CH2 over 10K), which seems to yield an offset of SNR between channels as his video shows at 8:20. OPA1612's datasheet (also refer to Fig. 7 about load vs THD+N) said in the last part that it is better to go with high gain with high impedance and high capacitance load (no circuit design for his gain stage).

What's your point. It is all about the difference between ch1 and ch2 which in case of an equal (or close to equal) signal results in almost no difference across the 10.3k resistance. There is virtually no 'load' to the input buffers.

Then my question is, is it justifiable that you balance out both signals with a trimmer resistor, expecting them to have same level of noise? IMO, it is quite clear that one channel of his null tester is ~1dB noiser.

Is there any other way than trimming the difference ? The buffer circuits at the input are 1x gain and have very low noise because of it. The difference in noise is then going into the null and will be audible when perfectly nulled. When this is at a level of -110dB what is there to worry about. ?

My -84dB refers to his source, which is the sansa clip mp3 player. I got one article on THD+N of one sansa clip model, which is -84dB. So I borrowed it. But, AFAIK, Sony WM1Z can only do -99 dB with 32ohm load. I don't know what would happen if the load is ramped up to 1.1K.

It is irrelevant whether the source has noise or not. Even when the source has 8 bits, -30dB noise with some music + heaps of distortion then this is absolutely of NO consequence and the null will still be possible at -110dB (and even lower as Ethan explained)
In fact you can use white noise as a source and null that.

The source quality is completely irrelevant as it is just a 'signal'. It is the null between that same source signal and via the cable under test to a similar load input. And then only the difference between them.

I won't ask so many questions if he was using a Line-level source with a more detailed specsheet and designed for high Z load.

I don't understand what you try to tell here.

 

[Blumlein 88]

If I may jump into this silly argument. Ethan is splitting and nulling over cables in real time I do believe (unless my memory is faulty). So in that case yes, you can have high noise, but it is the same noise on both inputs to the nulling apparatus and it will null out too.

As an illustration of the difference, I can create a file digitally and make two copies A, and B. They'd null out to nothing of course. I can add random generated noise to A, and generate random noise a second time and add to B. Those will only null out to the combined noise levels. If my noise level was -84 db in both cases the null will be limited to -81 db.

However, if I generate some random noise and add that same noise to both A and B, then again they will null out to absolutely nothing. This is more like what Ethan is doing with real time nulling. Other aspects may limit it from a complete infinite null, but getting -110 db or so is possible and noise levels don't prevent this in the scenario.

 

[Blumlein 88]

I moved on. Please check back on #539 and #540. I am working on a demonstration.

Which of the two scenarios I describe in my post will your demo fit into? Because I can tell you once we know that, we don't need to do the demonstration.

My post DIRECTLY gets at what you were saying in posts 539 and 540 which is what prompted me to post in the first place.

 

[Blumlein 88]

I am quoting solderdude's entire post from just a few posts upthread. It gets at exactly why there is a misunderstanding on this.

that means the null is not perfect. could be because of component tolerances between 2 channels.
In the null circuit this can be optimized.

In the end the null circuit has less components and does not rely on 2 potentially differing ADC channels and electronics.
You can see in Ethans video you can null cables so only noise remains.
My experience is the same. For nulling longer cables you need a delay line (cable of the same length) though.
When audio remains audible the null is not perfect.
I had to use a coarse and fine pot to get perfect nulls.

Without a shadow of a doubt the device Ethan has can obtain better nulls when one has a real-time device to test. DAC's are not real-time devices so that should be done with something like deltawave.

Deltawave is an excellent piece of software, but it has limits in fixing timing issues. Tiny timing issues can ruin a null. Even locking clocks and comparing two test runs is limiting due to noise. Ethan's real time nulling avoids all of this and in most circumstances (I'm tempted to say all) will get better, cleaner, deeper nulls than Deltawave. I think a basic misunderstanding is happening here about noise when it is real time nulled vs any version of recorded signals being compared that occurred at different times.

 

[solderdude]

I have formatting issue. So look at this one.

This is not true. Let's quote NwAvGuy here: Noise is cumulative. It only goes up even if you are to subtract A from A. In fact, doing subtraction of two very close numbers is highly problematic in error estimation. A better practice is actually dividing the two.

If we went through error propagation 101 again. Let's say subtracting 1.351+-0.002 with 1.349+-0.002, we will get 0.002+-0.0028 (or SNR=-1.5 dB). But for division you get 1.001+- 0.0021 (or SNR=22dB). You see the issue of subtracting close numbers now? Noise becomes larger than valid signal.

Besides, Ethan clearly noted in his video that nulling referred to the subtraction of all noise summed up. But obviously you can never sum a -84dB noise down to -110dB. That total noise (higher than -84dB) will determine your measurement's minimum detection limit (MDL). You cannot detect a valid difference smaller that limit (-84dB). Therefore you cannot subtract your result to get a valid difference smaller than that limit (-84dB).

You still don't seem to get it.

I have a signal with lots of noise. Lets say -84dB it is mixed with -110dB noise. Then the result is slightly worse than -84dB because the noise adds.
To calculate this you have to convert the dB to actual levels and then add the levels and convert back to dB.

-84dB (opposite 1V) = 63uV
-115dB = 1.8uV
it adds so converted to dB = -83.77dB so 0.23dB worse signal/noise ratio after the buffer.
The signal (after being buffered) passes through the test cable and through a second buffer.
So 64.48uV + 1.8uV = -83.53dB (in total 0.47dB worse).
Of course noise is random so sometimes it adds sometimes it cancels and everything in between.

The signal itself is the same + added noise.
When one cancels the 2 signals only the noise will differ and thus not cancel.
The -84dB noise is the same on both inputs thus cancels (it has the same waveform after all)
So the total noise that is different and thus produces a dif signal is 3.6uV... IF it were in the same phase.
But it is not. 1.8uV of it is random and the other 1.8uV of it is also random but has passed as a signal though the second buffer.

This means that noise of the null is higher than the added 1.8uV but below 3.16uV (because of the random nature)
Let's make it worst case. 3.16uV of noise is coming out (being different) and the signal with the -84dB noise is the same on both channels.
This means that when cancelled the 3.16uV is all that is left which is -110dB.
Of course the nulling opamp also has noise it adds as do the amplification stages.

But in the end regardless whether or not the test signal has -10dB noise or -90dB noise that noise is a signal which is common in both pads.
It does not matter how much noise is in the signal and how much distortion products. These are in phase in both A and B path.
Only the noise from the buffers is not in phase and won't null.

So yes.. no matter if the source file contains noise on a perfect null all of that is in phase and nulls. Only the differential noise is heard.
In the example the noise will thus be -110dB and NOT -83.5.

 

[Blumlein 88]

Well we can quote NwAvGuy here on a few things. This is his measurement of sansa clip+ back in 2011. http://nwavguy.blogspot.com/2011/02/sansa-clip-measured.html

In this article, he pointed out sansa clip+ has audible hiss.

So it is doubtful that Ethan's setup may not have the resolution to reproduce those minor audible difference on video, or does not reflect a good typical users scenario.

Also, while he mocked that $700 usd cable, I spotted some changes on his computer display. Whatever he got out from both directions, have an amptitude difference. Noise kicked in "faster" (no need to trim to a low difference) when he ignored direction marker of the cable.

Oh please.

You can compare 26:58 and 28:06 and see differences as large or larger than your two screen shots. The peaks and valley's over the scope are flickering around in the video naturally. The screenshots are meaningless in context. I don't know what it is you are trying to indicate, but you need to rethink this with the null tester.

 

[Blumlein 88]

You are aware that this is not numerically quantified, right?

I am straight on my indication. Even if you are nulling, you cannot say "ok, we get the null residue down to -110dB. That's our null noise in the measurement". This is rather obvious in terms of error calculation. As error in measurement is, for example (+-)5. You reversed it, it is still (+-)5. And you calculate (+-)5 add/subtract (+-)5 it always equals to (+-)10. There is a reason why it is called uncertainty and it follows Gaussian distribution.

Are you aware of common mode noise?

 

[EB1000]

Some people hear voices, and some people hear power cables audible differences... One thing is for sure, both are delusional...

 

[Blumlein 88]

Please check again.

So what is your point about all of this? I'm jumping in the middle I suppose I've not read every post of yours. Are you saying Ethan's null tester couldn't get a -110 db null if the source had -84 db noise? Are you saying Ethan mis-spoke when he said nulls were near -110 db which was the noise of the tester and should have said you'd expect nulls of -107 db instead? What is it your saying is wrong with Ethan's null tester or what is the point you have in regards to null testing? And try to answer more fully instead of coy short answers please.

 

[Blumlein 88]

You are aware that this is not numerically quantified, right? Just like some claim they can hear difference in cable.

Back to noise, not all types of noise can be nulled out. In common noise measurement, FFT is used to decompose noise into spectral component. But some noises cannot be decomposed, they look like what picture attachment shows. This noise has no phase, so you cannot reverse polarity and cancel it out in nulling. They only adds up even if you null them. The most ideal scenario is that you get it to be a constant output, but it is not going to be zero.

Therefore, nulling signal will only increase noise. If you want to compare cables like this, you have to be very careful.

If there is noise in the system of -110 db, and it is riding on both cables from a common signal, and the cables are only contributing -180 db of noise themselves, then you will get a result that looks within the margins of the gear like it did all null out. The cable noise of -180 db noise will add up to -177 db of noise, but the common noise of -110 db from the electronics is something that can be reversed in polarity and cancel out since the electronics as the source is common to both wire leads. Or if you have buffers which are each -110db them maybe you'll get -107 db. And this will be so even if the source for both leads is -84 db. Again I'm at a loss to see what your complaint or point or difference of opinion is.

 

[pkane]

In this process, only source noise in phase from A can be zeroed out, while noise not in phase will be added up.

There can't be out of phase noise coming out of the source -- there is just one output that's split outside the source.

 

[Deleted member 2944]

Chris,

The null test is A plus (negative) B.
It's a single source so the noise is correlated.

You're getting yourself all turned around and confused here.

Dave.

 

[pkane]

well, if you apply fft to white noise, you get white noise. That means white noise has no phase, so you cannot reverse polarity. In noise measurement, those baseline noise with strong fluctuations is such noise. They don’t go away in null operation.

White noise, pink noise, brown noise, blue noise etc. are all just examples of a signal that happens to have a random distribution. When you subtract a signal from itself you get a perfect null. This is true for any signal, including pure noise, music+noise or any other combination. FFT has nothing to do with this.

 

[Blumlein 88]

So Chris are you saying you don't believe Ethan got the result he is portraying in the video? He has a VU meter showing nothing at -20 db, and this with 80 db of gain applied. So he says the null residual is at least -100db and the gear has noise somewhere around -110 db. There is nothing wrong with any of what he says or does in this.

You seem all wrong-headed about real time noise, vs recorded noise and contributions from other sources.

 

[pkane]

only when you recorded it.

Waveform exists in the analog domain even if you don't record it. It can be split and inverted and summed without ever recording it or converting it into digital. That's what Ethan is doing. I don't understand your objections at all, I'm afraid.

regardless, have you done any error propagation that can reduce error? Or in practical term, an audio device that can add details to the source.

I don't understand the question.

 

[pkane]

with error “propagation” in mind, errors can only stack up. There is no way to correct it out. That’s conceptually incorrect to think error can be corrected out. So his null testing will definitely has a higher noise than his source.

He didn’t calculate error of his system properly

No. That's not how the null tester works. It doesn't care what signal is coming out of the source, noise, music, or other random signal. Whatever comes out is split into two.

Only noise and distortions introduced at or after the splitter can add up. Nothing before the splitter matters. There's no recording needed, no FFTs are part of the nulling process.

 

[Blumlein 88]

Unfortunately regardless of your mal conception of the situation you can plug in real sources, do the procedure, and get a result which contradicts your thinking. Therefore the error being propagated is the error in your conception of what is happening. That alone should give you pause to reconsider.

 

[pkane]

I am afraid Heisenberg won’t agree with you. White noise has a random vector, based on uncertainty principle, 1) you cannot measure both vector and position of an electron at the same time, 2) your measurement of position will disturb its vector.

So you won’t get a reversible waveform unless you record it.

So you are worried about noise due to quantum uncertainty principle, is that it?

 

[Blumlein 88]

well impedance and voltage are both complex numbers with imaginary numbers, which can sometimes be zero.

thing is, why are you obsessed with how people interpret circuit like it was 100 years ago.

one thing here is the VU meter. It’s designed for 600 ohms source output impedance, and that’s a fairly old standard. Modern solid state audio devices have a typical output impedance of 100 ohms, with actual value ranging from 50-300 across manufacturer. Jumping from 100 ohms to 600 ohms, with modern recording interfaces and long cables, you will get -7dB signal loss in 10 khz compared with nothing before the jump.

He has a driver circuit for the VU meter. It isn't in series with the rest of the circuit. Sheesh.

As for dbV or dbu, if you properly compare ratios it doesn't matter. 80 db of gain is 80 db of gain whether you report the final result as dbV or dbu.