• WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

How flat is flat?

daftcombo

Major Contributor
Forum Donor
Joined
Feb 5, 2019
Messages
3,687
Likes
4,068
Hello,

Let's take two speakers A and B identical in all aspects but flatness of their frequency response.

Let's say, for instance, that A is flat +/-0.5 dB and B flat +/- 1 dB.
Is the difference audible?

Of course, it's not possible to find two speakers identical in all aspects but FR. But it could be possible to test that by taking one speaker flat +/- 0.5 dB, EQ it (in an evil way) so that it's only flat -/+1 dB afterwards, and do an ABX.

What would be the limit of +/- x dB under which x or x-epsilon can't be ABXed?
 

sergeauckland

Major Contributor
Forum Donor
Joined
Mar 16, 2016
Messages
3,440
Likes
9,100
Location
Suffolk UK
Hello,

Let's take two speakers A and B identical in all aspects but flatness of their frequency response.

Let's say, for instance, that A is flat +/-0.5 dB and B flat +/- 1 dB.
Is the difference audible?

Of course, it's not possible to find two speakers identical in all aspects but FR. But it could be possible to test that by taking one speaker flat +/- 0.5 dB, EQ it (in an evil way) so that it's only flat -/+1 dB afterwards, and do an ABX.

What would be the limit of +/- x dB under which x or x-epsilon can't be ABXed?
I doubt whether that sort of difference, +/- 0.5dB or +/- 1dB would be audible on programme material, or even on broadband noise. It may depend on just where the variations are, and whether they are broad or narrow. If broad, then on a rapid AB switch, they could be audible on noise, if narrow, very doubtful. Without rapid AB switching, i.e. just listening to one loudspeaker, then listening to the other, especially if there is an appreciable time lapse between switching, these sorts of differences are just not audible.

I don't know what the limit is where differences become clearly audible, as it depends on the type of programme material, and most importantly, on the accuracy of level matching. One problem with level matching loudspeakers, is that at what frequency is the matching done? All loudspeakers have variations in frequency response and sensitivity, so trying to match at one frequency could leave a considerable difference elsewhere which would be noticeable on an AB test even if the frequency response wasn't that different.

My own personal criterion is that if the frequency response is within 3dB (i.e. +/-1.5dB) over the audio band depending of course on the LF capabilities of the loudspeaker, that's good enough. I've equalised my own to be +/- 1dB above 200Hz as anything closer than that was just too much bother.

S.
 

Fluffy

Addicted to Fun and Learning
Joined
Sep 14, 2019
Messages
856
Likes
1,423
Rfegwbz.jpg
#sorrynotsorry

Now to the point:

+/-0.5 dB is extremely hard to achieve, and even +/- 1 dB is rare. I think that if you have speakers that are actually +/- 1 dB flat you're pretty much golden.

You can't really directly compare those two levels of flatness, because overall sound will depend more on where are the biggest variants in frequency. If one speaker is +1 db in the bass range and -1 in the treble, and the other one is the opposite, they are both technically +/- 1 dB flat, but will sound different.
 

solderdude

Grand Contributor
Joined
Jul 21, 2018
Messages
15,891
Likes
35,912
Location
The Neitherlands
Hello,

Let's take two speakers A and B identical in all aspects but flatness of their frequency response.

Let's say, for instance, that A is flat +/-0.5 dB and B flat +/- 1 dB.
Is the difference audible?

Of course, it's not possible to find two speakers identical in all aspects but FR. But it could be possible to test that by taking one speaker flat +/- 0.5 dB, EQ it (in an evil way) so that it's only flat -/+1 dB afterwards, and do an ABX.

What would be the limit of +/- x dB under which x or x-epsilon can't be ABXed?

If you want to test... simply use a parametric equalizer and randomly shift some of the sliders up and down 0.5dB to 1dB and switch (blind) between 'flat' and 'EQ'.
This works well even with wonky speakers and headphones as basically you only want to find out if such small differences are audible.
 
Last edited:

pozz

Слава Україні
Forum Donor
Editor
Joined
May 21, 2019
Messages
4,036
Likes
6,827
You probably have random irregularities in mind. But Q is important. Broadband 0.5dB trends in the midrange and highs are audible.
 

PaulD

Senior Member
Joined
Jul 22, 2018
Messages
453
Likes
1,341
Location
Other
You probably have random irregularities in mind. But Q is important. Broadband 0.5dB trends in the midrange and highs are audible.
Pozz is absolutely correct. I have first hand experience of a couple of people noticing a 0.1dB broadband change. Could not identify what it was, but noticed that it was "different" than before. Of course they knew the replay system very well (a studio), and it was blind (they did not know anything had changed, it was not a "test").
 

QMuse

Major Contributor
Joined
Feb 20, 2020
Messages
3,124
Likes
2,785
View attachment 67296
#sorrynotsorry

Now to the point:

+/-0.5 dB is extremely hard to achieve, and even +/- 1 dB is rare. I think that if you have speakers that are actually +/- 1 dB flat you're pretty much golden.

Let's not forget what kind of mess room does to a speaker below 300Hz, and good luck with EQ-ing that to such linearity over the listening area. ;)
 
Last edited:

Speedskater

Major Contributor
Joined
Mar 5, 2016
Messages
1,610
Likes
1,333
Location
Cleveland, Ohio USA
With a skilled listener and very demanding A/BX testing, differences smaller than ±0.5 dB should be detectable. But real world, real music is another story.
 

edechamps

Addicted to Fun and Learning
Forum Donor
Joined
Nov 21, 2018
Messages
910
Likes
3,620
Location
London, United Kingdom
Let's take two speakers A and B identical in all aspects but flatness of their frequency response.
Let's say, for instance, that A is flat +/-0.5 dB and B flat +/- 1 dB.
Is the difference audible?

It depends on many factors, some of which are discussed in this famous paper by Toole & Olive. Most importantly, it depends on Q of the variations as well as the content being played, so a simple number like "+/-0.5 dB" is not enough to conclude anything. Notably, deviations become more audible as Q decreases (i.e. broadband variations are more audible than sharp peaks) and content resembles broadband noise (as opposed to, say, clicks). In the most extreme case, the study has shown listeners can detect a ~0.25 dB deviation with Q=1, on pink noise, using loudspeakers in a room.

A spectral "tilt" or "slope", which is basically a deviation of even lower Q, is the most audible deviation there is and can be noticed even if the slope is very small. However, do note that this is all about detection thresholds, not preference, and in the case of broadband deviations one could argue we can become "adapted" to the sound and "see through" the anomalies.
 
OP
daftcombo

daftcombo

Major Contributor
Forum Donor
Joined
Feb 5, 2019
Messages
3,687
Likes
4,068
+/-0.5 dB is extremely hard to achieve, and even +/- 1 dB is rare. I think that if you have speakers that are actually +/- 1 dB flat you're pretty much golden.
Some DIYers pick flat drivers and use digital crossovers with 300dB/octave steep slopes (FIR). The baffle step once corrected (not even needed for in-wall mount), it is definitely possible to have very flat speakers.

So my point is: how useful is that? Can a loudspeaker "not so flat" be audibly similar? I had high Q in mind, as you guessed. Low Q is supposed to be easily corrected with EQ.
 

sergeauckland

Major Contributor
Forum Donor
Joined
Mar 16, 2016
Messages
3,440
Likes
9,100
Location
Suffolk UK
Some DIYers pick flat drivers and use digital crossovers with 300dB/octave steep slopes (FIR). The baffle step once corrected (not even needed for in-wall mount), it is definitely possible to have very flat speakers.

So my point is: how useful is that? Can a loudspeaker "not so flat" be audibly similar? I had high Q in mind, as you guessed. Low Q is supposed to be easily corrected with EQ.
I don't know of any driver that's flat to 1dB or better without external EQ. Those loudspeakers that have very flat responses use DSP to correct for the limitations of the drivers. As to the slope of the crossover, that doesn't in itself flatten the driver, but a steep slope will avoid the driver being given anything to do outside its passband, which helps with distortion, and may in the limit help suppress sharp High Q resonances outside their passband which could affect the overall frequency response. 300dB/Octave does seem somewhat excessive though. I use 48dB/octave and 24dB/octave is more common, whilst passive crossovers are anywhere between 6and 24dB/octave, anything over 24dB/octave being impractical with passive components and even with active filters, is about the limit without DSP.

If an uncorrected loudspeaker has multiple high Q resonances, that's indicative of poor drivers, but won't necessarily sound as bad as a graph might suggest given that each peak will have little energy associated with it due to the High Q.

S.
 

GelbeMusik

Senior Member
Joined
Apr 19, 2020
Messages
445
Likes
290
It depends on many factors, some of which are discussed in this famous paper by Toole & Olive.

My AES membership still has to start one day. I won't pay the extra fee for wisdom, which was discussed for centuries already. The terminology seems to hinder handy results. Why is it, that people speak of "Q" with simple amplitude ondulations? The technique, as far as I could conclude from the public narrative: take a signal, split it, run one path through a so-and-so Q-filter, sum up the paths with individual factors after, so that some bump / dip of so-and-so high is achieved. Right?

This isn't a good model of speaker frequency dependent amplitude response ondulations. I don't know if it would make a good stimulus for the human hearing either.

Again, I don't know if the paper discusses this. If not, I think we should take the results with a tablespoon of salt. If though, would sombody give a short summary?
 
  • Like
Reactions: Zvu

sergeauckland

Major Contributor
Forum Donor
Joined
Mar 16, 2016
Messages
3,440
Likes
9,100
Location
Suffolk UK
This one between 400 Hz and 1.6 kHz for instance:
http://www.troelsgravesen.dk/ATC-SM75-150.htm
It actually isn't that flat, some 2 dBs variation between those frequencies, but pretty good, although it's a fairly restricted range, only 2 octaves. Between 300-3kHz, which is where I would expect this driver to be used, it has something like a 4dB variation. Still very good for an unequalised radiator, and one of the best mid-range drivers available, but rather more than 1dB.

S.
 

edechamps

Addicted to Fun and Learning
Forum Donor
Joined
Nov 21, 2018
Messages
910
Likes
3,620
Location
London, United Kingdom
Why is it, that people speak of "Q" with simple amplitude ondulations?

Q is a perfectly acceptable way to characterize the "bandwidth" of an FR peak/dip (i.e. ondulation). It maps nicely to the physics of resonances and the parameters of a peak EQ filter. It's also quite specific and unambiguous. What would you have used instead?

The technique, as far as I could conclude from the public narrative: take a signal, split it, run one path through a so-and-so Q-filter, sum up the paths with individual factors after, so that some bump / dip of so-and-so high is achieved. Right?

Right. And in some tests they also add delay to the filtered signal (i.e. a delayed resonance) and see what happens.

This isn't a good model of speaker frequency dependent amplitude response ondulations. I don't know if it would make a good stimulus for the human hearing either.

Why? If you're worried about the time domain, that's discussed in the paper, which includes an experiment that purports to show that what we hear in these tests are the bump in the frequency domain, not the associated ringing in the time domain.
 

GelbeMusik

Senior Member
Joined
Apr 19, 2020
Messages
445
Likes
290
Q is a perfectly acceptable way to characterize the "bandwidth" of an FR peak/dip (i.e. ondulation). It maps nicely to the physics of resonances and the parameters of a peak EQ filter. It's also quite specific and unambiguous. What would you have used instead?

What about deviations with a very wide bandwith, say from 1kHz to 4kHz, a shift upwards by a constant 1..2dB? That should for sure be distracting to "critical listeners". It could not be modeled by a single Q "resonance". I'm not concerned about the time domain.

The Olive metric doesn't take such deviations into account. It focusses very much on narrow deviations, but neglects trends completely and doesn't care explicitly about wide-band deviations. Maybe it is because of the "Q"-model.

At least the "Q"-model isn't easily adopted to the case of multiple dips n' peaks, may they be alternating or accumulate with the same sign somewhere on the frequency axis--so far my assumptions as a non-member of AES.
 

edechamps

Addicted to Fun and Learning
Forum Donor
Joined
Nov 21, 2018
Messages
910
Likes
3,620
Location
London, United Kingdom
What about deviations with a very wide bandwith, say from 1kHz to 4kHz, a shift upwards by a constant 1..2dB?

Very wide bandwidth just means very low Q. I'm not sure what your point is.

The Olive metric doesn't take such deviations into account. It focusses very much on narrow deviations, but neglects trends completely and doesn't care explicitly about wide-band deviations. Maybe it is because of the "Q"-model.

The Olive model does not use filters with Q anywhere (not in its development, neither in the model itself). In fact, it kinda does the opposite because NBD uses arbitrary fixed octave bands with discontinuities between them - the exact opposite of what a smooth "Q-based" filter does. (That's probably a bad thing, by the way.)

At least the "Q"-model isn't easily adopted to the case of multiple dips n' peaks, may they be alternating or accumulate with the same sign somewhere on the frequency axis

But of course it is, you just need multiple filters. Have you ever used a graphical or parametric EQ? It's just a series of peak filters, each with their own Q, cascaded together, and with enough of these filters you can achieve pretty much any frequency response that way. And conversely, you can interpret any frequency response (such as a speaker frequency response) as a cascade of peak filters. Especially if it's minimum phase.

Now, it is true that the study only used one filter, but that's because the goal is to understand how we perceive a single resonance (i.e. a single peak). Keep in mind that study is investigating the audibility of resonances across a number of dimensions - frequency, Q, level, delay, content, listening method (speakers in room, speakers in anechoic chamber, headphones). That's a ton of variables already. Adding more filters with more parameters on top of that would likely have made the study impractical.
 

GelbeMusik

Senior Member
Joined
Apr 19, 2020
Messages
445
Likes
290
Have you ever used a graphical or parametric EQ?

C'mon.

It's just a series of peak filters,
...
Now, it is true that the study only used one filter, but ...

I'm o/k with that. But, as You said: "but", it is a long way from there to understand the effect of pretty nice ondulations or even accumulating peaks, for instance the infamous "bathtub" or "smiley" frequency response. The "clearing" peak around 5kHertz and so on.

Anecdote 11: listened to a demonstration of "high end audio", the sound (!) on YOUTUBE :facepalm: but, again it was still revealing :cool: . It all sounded the same: wrong. Too "clear". Tja ...

With the Olive rating I'm not o/k. But that's another topic.
 

tuga

Major Contributor
Joined
Feb 5, 2020
Messages
3,984
Likes
4,281
Location
Oxford, England
Adding more filters with more parameters on top of that would likely have made the study impractical.

A bit like using a pair of speakers in A/B blind tests... Trade effectiveness for practicality.
 
Last edited:

aarons915

Addicted to Fun and Learning
Forum Donor
Joined
Oct 20, 2019
Messages
685
Likes
1,140
Location
Chicago, IL
I take this question as how flat is flat (enough). Using the LS50's as an example, their listening window in the Soundstage measurements fits in a 3 db window and the PEQ filters I use flattens it to a 1.5 db window. I would say the difference is pretty subtle when comparing my filters vs No EQ so based on that I would say if you can get your listening window within a 2db window (with a speaker with good directivity) you're in good shape.
 
Top Bottom