How do you judge an EQ?

[Zaphod (giancarlo)]

Infact big ears mixers have good gear

and a big imagination especially when they look at the tools they use.

no, simply good ears. Remember that only today we have an accurate value on our monitors. I've sampled a lot of hardware gear, and, believe me, if something is labelled 1dB, probably it is 1.1 or 0.9

[Zaphod (giancarlo)]

i absolutely agree, i don't get his argument. everything is degrading to the sound with further processing, analog or digital. what people describe as a destruction of transients could be in fact a way neutral behavior regarding transients and what enhances them could be through actually adding something to the sound, but isn't that degrading?
the example is also not real-life imho. i don't use 10 different eqs on a single source and if you would do that with an analog eq you would also degrade the sound, maybe way more than with a digital eq...chances could be high that people would like the analog chain just because of the heavy degredation.

The point is that you haven't the same S/N on every hardware, and you haven't the same unwanted distortion on every software. My point is: in software there is something that is better for transients and something not. I agree with bmanic here. Imho an eq is good when transients are preserved better. Probably other people want better precision/accuracy (if I ask 1 db, I WANT 1dB), other prefer a better behaviour around the Nyquist frequency. This topic is exactly on this question. I see everyone is answering to the same question in a different way.

[defjamm]

Infact big ears mixers have good gear

and a big imagination especially when they look at the tools they use.

no, simply good ears.

good ears don't help with the human brain and what you see with your eyes and more important what you believe how it should sound. believing is really strong, not just in religion, it's the same with audio.

[Barbarossa]

Imho an eq is good when transients are preserved better.

Any foundations?
Following this argument, linear phase EQs should sound always best. My experience is, that isn't the case. Famous analog EQs are just ordinary phase-minimum filters. And that means, they smear transients.
Our ear isn't a logical machine. A certain amount of noise, distortion and other artifacts we find pleasing. Why not smeared transients?
The difference between a good and a not so good EQ must therefore lie somewhere else.

My guess is, the filter coefficients determining the filter's response are very important and here do some EQs offer more and others offer less "musical" shapes.
But the real clue that makes the last 10% of the magic sound of high end analogue gear, lies in the non-linearities of analog circuits. I think this is the main reason, why top end analog gear hasn't been reached "quality" wise by digital emulations.

Christian, it's high the time to become famous and develop a closed theory about non-linear time invariant systems...

[defjamm]

My point is: in software there is something that is better for transients and something not. I agree with bmanic here. Imho an eq is good when transients are preserved better.

i'm actually with you on this one, there are differences. but differences have their reason and i want to understand what i'm hearing and why. that's where science comes into play and becomes very important. i don't want christian or any other developer to tell me how i should equalize something but i enjoy their input how somethings works from the technical point of view. it helps me to learn why i'm hearing what i'm hearing.
regarding preserved transients...i don't know it preserved is the right word, maybe enhancing describes better what some devices do and why we humans like it.
if you're used to enhanced transients for 20 years and than hear something that doesn't enhance them that strong it will sound strange and wrong. but the reason is imho not the format, it's the implementation. if you want to copy a with b you need to know what a exactly does, that's where the science comes into play(i like this sentence therefore i'm repeating it ).

[martian]

can those with differing opinions and findings to me post some more examples please. if there really is something relevent im missing that i can hear through my humble monitoring set-up (mackies, hd600s, delta 44, acoustic treatment), then i care 100%!

so zaphod and bmanic, can you objectively demonstrate your transient issue? to clarify its digital biquad min phase eq comparison that is the subject matter here for me.

furthermore i am of course open to the fact you may have much better hearing and monitoring, but i still wont care if i cant hear it. more power to you if you can which i wont doubt if the tests are objective.

thanks,

[Christian Budde]

But the real clue that makes the last 10% of the magic sound of high end analogue gear, lies in the non-linearities of analog circuits. I think this is the main reason, why top end analog gear hasn't been reached "quality" wise by digital emulations.

The 'analog' mode in my EQ's already cove some basic non-linearities. At least those which can't be eliminated in analog circuit design. They can be reduced, but never eliminated in completely.

Christian, it's high the time to become famous and develop a closed theory about non-linear time invariant systems...

That's no problem, but I simply don't have gear to measure and reproduce.

[living sounds]

So, could Christian or someone else explain what happens to those poor transients? I'm absolutely sure that whatever signal I EQ with (for instance, there are others, lot's and lot's) the Cambridge loses definition, sounds smeared - and boomy, thin, screaming (depending on what frequency you boost/cut). Hydratone, the UAD-1 Neve or the A0 parametric EQ don't do that. On the contrary, you can make it sound big, bright, heavy whatever, but it won't be destroyed like that (A0 is not artifact free, but it has THAT punch). You can dial in too much or too little, but in general the sound will be different, but not worse (in terms of transients). So how come nobody got this right until recently?

[Eiosis]

Just make this little test by yourself. Put 0.1 dB 10 times with a bad plug.
Today we use lot of plugs in long chains, pay not much attention on the single degradation, but at the end of the chain you have a big mess.

I did! I compared 10 times 1dB against 1 time 0.1dB. with te result I did a double blind test and I figured out that at least I don't hear a difference. [...]
If you do this test on your own, you have to be sure that you remain the bandwidth constant. If you have an EQ such as AirEQ, your parameter mapping isn't linear. This will in fact sound very different than.

I think that what Zaphod meant is not exactly the same thing.

He probably wanted to say that if you process ten times a single track, for example with a bell filter at 1kHz, Q=1 and Gain=0.1 db, you should have an almost untouched transient with the bell filtering added result, i.e. a 1dB bell boost @ 1 kHz, Q=1, which must not affects the original transient a lot, because the added energy of a 1khz 1dB Q1 bell is rather low. (i.e. 10 times 0.1 dB)

Some plugins could affect the original transient more than expected, and I can be wrong, but I think that Zaphod's test purpose was to exagerate the way that the original transient is affected by an EQ, by reapeating 10 times the processing of this plugin, and not to see if 0.1dB x 10 = 1 dB of boost or 1dB x 10 = 10 dB of boost.

... can you objectively demonstrate your transient issue?

...what happens to those poor transients?

I tried this test like described with two digital EQs,
i.e. : 10 times a bell boost @1khz, Q=1, G=0.1 dB.

Here is the results : (all my apologies for the big sizes )

RED : original transient
BLACK : processed transient


First screenshot :
On the upper response, the original red transient and processed black transient are merged.
For the lower one, they are not merged at all.




Second screenshot :
An amplitude zoom version of the first screenshot.


We can see that for the upper impulse response, the original transient is still here,
and that for the lower IR, the transient is blurred a lot after 10 processings.

The frequency result is about the same, i.e : a bell boost @ 1Khz, 1dB (10 x 0.1dB) Q=1, sampled @44.1 kHz 32bits.

I don't know if we can say that one equalizer used for this test is better than the other,
but it would be very interesting to do the same test with some analog EQs.
The test would be more complicated with analog EQs because it must take into account the whole audio signal path
(D/A conversion - out amp - audio cable - EQ - audio cable - in preamp - A/D conversion).

We should also think about the fact that when we use an external analog equalizer for mixing purposes on a digital system, we have to deal with the imperfections of the whole audio signal path, which sometimes can be worse than just blurred transients.

An analog EQ alone is probably of the best quality, but an analog EQ plus low-end A/D and D/A converters may not be very good.


Fabrice,
Eliosound

[defjamm]

great information! you won't tell us which eqs were used for your examples? regarding transients, distortion should also play some role or am i wrong here?

A0


hydratone


gliss

[Christian Budde]

Just make this little test by yourself. Put 0.1 dB 10 times with a bad plug.
Today we use lot of plugs in long chains, pay not much attention on the single degradation, but at the end of the chain you have a big mess.

[..]

I increased the gain by a factor of 10, oopsi.

I think that what Zaphod meant is not exactly the same thing.
[..]

Ok, got the point.

Some plugins could affect the original transient more than expected, and I can be wrong, but I think that Zaphod's test purpose was to exagerate the way that the original transient is affected by an EQ, by reapeating 10 times the processing of this plugin[..].

... can you objectively demonstrate your transient issue?

...what happens to those poor transients?

Now I think there is maybe a lack of definition here (or I can be totally wrong here). The thing you describe is the effect of limit cycles. These are especially high for the clinical test you provide. Here's some description of limit cycles:

Normal recursive filters (i mean not those based on convolution) always have a feedback loop. This is the same for digital and analog EQs. But in contrast to the analog domain, the digital domain is discrete, that means the feedback data is quantized. This data is put again to the input and everything starts over nd over again. especially if you have a dirac impulse (like in your example) you soon will have very huge quantization errors. Even if you use floating points, especially for poor numerical implementations.
Example: Let's say you have only 3digits and you have to add the hypothetical signal 8.33768 and -8.33521. The correct answer would be: 0.00247 or in our 3 digits: 2.47 * 10^-3. If you truncate before adding this signals you get: 8.33768->8.33 and -8.33521->-8.33 In this case the result is 0.
If you feed the filter with a wrong feedback, you sometimes might produce cycles. The system is trying to reduce the feedback to zero (because of the coefficients), but this fails, because it is always rounded up.
You can hear those limit cycles very good, if you quantize the state variables (aka. feedback) to 8Bit or so.

An indicator for limit cycles here is the fact, that a high frequency ringing is shown

But I may be on the wrong path here as well.

Another explanation would be, that an (linear phase) oversampling filter is used. These filters are very steep cutoff filters near nyquist frequency. Depending on the frequency, there should be also some ringing here. But in that case you would also have it ringing after processing the signal only one time.

It would be kind, if you can tell us, what gear you've used.

great information! you won't tell us which eqs were used for your examples? regarding transients, distortion should also play some role or am i wrong here?

A0
[..]

hydratone
[..]

gliss
[..]

An interpretation of the distortion plots isn't always easy, especially if plugs introduce unreported latency. In that case something like shown for the A0 and hydratone happens. A plugin with consequent use of 32bit data (and more for internal processing), should get easily below -120dB (and more)
Anyway, these charts prove, that there is no harmonic distortion in this plugins (a second line at 2kHz would indicate this)

You can dial in too much or too little, but in general the sound will be different, but not worse (in terms of transients). So how come nobody got this right until recently?

I think, that there is no right and wrong. It's maybe a matter of taste and/or a lack of definition. If you're comparing apples with oranges (because maybe the definitions between two EQs don't match), you'll end up in chaos. So it's important to keep some features constant, while changing others. I have only started my journey yet, but I'll keep on investigation. One of the items on my list is to do a worldwide doubleblind tests, but it's hard to do, because everyone prefers other music, EQs, and so on. I'll start with an objective comparison between minimum and linear phase. With untrained subjects I could verify the literature and prove that no one was able to hear a difference. In that case the music and the EQ setting was fixed. My idea is to make a tool which does AB testing on arbitrary material. You'll setup your EQ setting in any mode and then the data is rendered. This data will be feed into a double blind test. The results are encrypted and can be mailed to me. After two month I will make an evaluation tool available, so that everyone can see, how he performed...

[Zaphod (giancarlo)]

Thank you eliosound. With a good example and a good english you explained very well what I meant.
That 'mistery eq' comes from a project that is trying to resolve exactly this issue. Bmanic (he is not linked to me directly) used it for asking to this community what you think about, I'm happy he did it (and it was a surprise, not programmed), because in my opinion this is a weak point for many digital plugs (not of all anyway)

for christian: these irs are 'intentionally' limited, but they could be enlarged as you want. Think to this mistery eq as an example. Irs came (I don't know if mistery1 or mistery2) from a top high-end unit, the Avalon AD 2055 (I suppose this is a legendary unit from Avalon for a lot of engeneers). But the point is not 'from wich unit'. The point is 'are transient preserved better than xxx or not' and if yes 'why, what is the reason' and 'who cares'

I'll shut up again now
byebye

------------
remember: resistance is useless

[jah1234]

- The most common error with convolution filters are, that the IR is too short. This will result in a different frequency response (compared to the original), especially for a high Q (narrow band). Additionally if long IRs are truncated without proper windowing, you get errors by hard cutting. These errors can be found in Tritone's products as well as in the Sintefex IRs. They result in ripple (see "Mystery 2" - may also result from a non-biquad realisation = multiple filter stages).

I have a spare time project coding a rather simple convolution EQ in order to learn DSP basics, and obviously I've also noted the impact IR size and windowing has on sound. During development I've also noticed minor flaws in other convolution EQs, such as Tritone products (some IRs in Tritone products simply seem to be ordered wrong!!).

For 44.1 khz sample rate, what is in your opinion the minimum IR size for a convolution based EQ? I didn't really follow you on the resolution discussion.

Thanks for your valuable insights, comments and plugins!

[Barbarossa]

But the real clue that makes the last 10% of the magic sound of high end analogue gear, lies in the non-linearities of analog circuits. I think this is the main reason, why top end analog gear hasn't been reached "quality" wise by digital emulations.

The 'analog' mode in my EQ's already cove some basic non-linearities. At least those which can't be eliminated in analog circuit design. They can be reduced, but never eliminated in completely.

But i mean the authentic reproduction of non-linearities out of a theory, not producing non-linearities, that just sound as similar as possible.

Christian, it's high the time to become famous and develop a closed theory about non-linear time invariant systems...

That's no problem, but I simply don't have gear to measure and reproduce.

I thought including non-linearities are only a product of observation and implementing something, that comes close to the measurements, but i thought there's no closed theory available, i.e. how a tube creates harmonics, like we have this theory for LTI-systems.
If this theory for non-linear systems would exist, then the whole problem of modelling analog gear would be solved already, wouldn't it?

[Christian Budde]

I thought including non-linearities are only a product of observation and implementing something, that comes close to the measurements, but i thought there's no closed theory available, i.e. how a tube creates harmonics, like we have this theory for LTI-systems.
If this theory for non-linear systems would exist, then the whole problem of modelling analog gear would be solved already, wouldn't it?

It's similar like taking IRs and doing convolution. Every recursive filter has an infinite IR, so you have to truncate it at some point. The position limits the quality of reproduction. So an IR of 256sample will have a frequency resolution of 256/44100 = 172.2Hz. So (for example) between 0-172.2 Hz interpolation takes place. If original EQ has a notch at 50Hz, you simply won't get it...
With nonlinearities it's nearly the same, there are several options available, to reproduce these nonlinearities. With these methods (and a fast computer) you can come close. Very close, I mean theoretically infinitive close, but the developer has to decide, how clos he wants to go.
The Tritone guys obviously decided that IRs at a maximum of 512 samples are enough (=86Hz resolution. Ok, accidentially some IRs got shuffled, but hey).
But back to nonlinearities. There are several methods known for reproduction. One for example is the patented dynamic convolution. This is one extreme, because it deals with simply stacking impulse responses for different levels. With this technique, you have the possibility to cover short time nonlinearities. For others it eats to much CPU and you maybe emulate things you don't want to. Also you may have problems with capturing the IRs. A similar technique has been described, to cover only the harmonic distortion (and not quantisation for example). You can come closer to the real sound with less computation, but again "capturing" the original is a bit hard.
On the other extreme you have al kinds of physical modelation, including various numerical methods. That's how I implemented the nonlinearities of my hypothetical analog EQ (i never build it in real hardware). In that case I also decided only to cover the strongest non-linearity, others I don't care about. As I mentioned above it's a matter of how exact you want to emulate the sound.
Since there is no magic in analog EQs, it's possible to make them that close, that no human ear is able to determine a difference, but at least there is the psychological aspect. No one will believe the fact (especially those who own expensive analog gear).

Independent double-blind tests might prove that, but...

Kind regards,

Christian

P.S.: Sleeping a night, I come to the conclusion, that the "smeared" IR must be the result of a linear phase oversampling filter. So it should be the first question to ask, if the difference of linear phase and minimum phase can be percieved. Obviously it can be perceived if you have a look at the plot. The psychoacoustics says, that it should get mask (pre-masking).