For DIYers designing completed loudspeakers, it can be quite challenging to get a low noise floor as low as 10dB, at least in my current part of the world - I have the sounds of birds calling, rustling leaves and modern machinery to contend with. If designing transducers professionally, or measuring the noise of inherently low noise things, DIYers can build their own dedicated hemianechoic chamber with wedges in their home, but this is not something I’ve (yet) done.🤓🥷
On the other hand. Measuring indoors has its own challenges; to block out the room reflection, we have to use gates. Depending on the measurement distance this is typically 5-10ms for a sweep, which gives a measurement resolution of 200-100Hz.
So even though REW shows a nice smooth graph, for instance, between 500Hz and 2KHz, but 1500Hz/100Hz= equivalent to 15 points of data, for a musical span of 2 octaves.
@lrisbo
N.B. When displaying the distortion, for the sweep REW shows the fundamental with a 1/24 octave filter, giving it a smoothed appearance. This is fixed, and there is no option to turn it off.
On the other hand. Measuring indoors has its own challenges; to block out the room reflection, we have to use gates. Depending on the measurement distance this is typically 5-10ms for a sweep, which gives a measurement resolution of 200-100Hz.
So even though REW shows a nice smooth graph, for instance, between 500Hz and 2KHz, but 1500Hz/100Hz= equivalent to 15 points of data, for a musical span of 2 octaves.
@lrisbo
N.B. When displaying the distortion, for the sweep REW shows the fundamental with a 1/24 octave filter, giving it a smoothed appearance. This is fixed, and there is no option to turn it off.
Last edited:
No, the smoothed appearance comes from REW.
If you just look at other programs with 1/24 filtering it doesn't have that smoothed look.
VituixCAD has this smoothed look as well, even at a lower octave filtering.
It reminds me a lot when Microsoft started to do the same thing in Excel.
Not that it matters because we are talking about super baby tiny scribbles that gave no resemblance to any practical information.
Hi, just for designing speakers.....why would we ever need such a low noise floor?
(Truly curious, not trying to be confrontational.)
a good 1/2” measurment mic has self noise around 17dB so not much point in trying to make it quieter.
My point is that with proper averaging and signal processing we can resolve details and harmonics up to 30-40dB below the noise floor
My point is that with proper averaging and signal processing we can resolve details and harmonics up to 30-40dB below the noise floor
that’s clever: removes the room from non anechoic measurements. Otherwise a long reverb tail of a room can affect harmonic distortion reading when using the chirp (higher harmonic responses bleed into lower harmonics). Of course, a very slow chirp gives more time separation of the harmonics
So we have to diskuss that with John Mulcahy, or we have to explore Farina exp sine sweep with ARTA.
Given the fact that REW applies a smoothing to 1/24 oct no matter what, a longer data acquisition will not help frequency resolution. @tktran303 explored the length and found no benefit for higher than 256K samples. He published that well.
Taking the spared time for 8 repetitions is a more useful investment, as this brings down the noise.
I don't see any need to discuss this smoothing any further, because STEPS 1/48 oct is not really better (see #66). A STEPS 1/48 oct measurement diagram of the whole 20Hz-20khz range visually pretends to give more resolution. But to evaluate more detail one has to stretch/zoom in a small frequency range. That is the way i do it. But this is not publishable unless one finds something interesting.
If all the ""distortion"" diagrams made with REW Farina over the years here on diyAudio had had the quality that @tktranh offers here today...
"REW (Farina log sine) Sweep, settings 256K samples x8 repetitions: time taken: 45 seconds:" is the way to publish the distortion of your project here on diyAudio.
best regards,
Bernd
The resolution is set by the smoothing so longer sweeps will of obviously not change that. What a longer sweep can do is to lower the noise floor and give better separation of the harmonics.
the question is why the noise floor on the 8x averaged plot is so high? A 8x longer chirp gives the same noise advantage plus better separation of the harmonics
A stepped sine samples the response at discrete frequencies and this can give aliasing. Samples at 1/48 octave is not the same as smoothed to 1/48 octave. The sampling theorem requires us to band limit before sampling. That is the challenge with the stepped sine.
The chirp method finds the impulse response and we can do true smoothing of the spectrum (which lowers the noise floor if done right), ie avoiding aliasing.
However, the stepped sine some offers huge noise immunity due to the long averaging. But an extremely long chirp with the duration of the stepped sweep should have similar noise suppression. It is of course the computer memory and processing power that can limit how long chirps can be.
All in all, the chirp method seems to be the winner if done right.
REW manual states that
"The Distortion graph shows the measurement's fundamental (the linear part of its response), its harmonic distortion components up to the ninth harmonic, Total Harmonic Distortion (THD) and the level of the noise floor, which is captured before the measurement starts.
The plots are derived either from analysis of the impulse response or from stepped sine measurements.
Reference:
https://www.roomeqwizard.com/help/help_en-GB/html/graph_distortion.html#top
Lars, when say "chirp", are you referring a signal that is different signal from REW's log sine sweep?
@mark100
That is a great question and thanks for raising it. But I was was hoping someone would raise it. I'm not sure I can answer this one easily, but I was planning to summarize all my findings in my original post.
So far, going through this process clarified some things.
In the context of DIYers ie. hobbyist / non professionals- the measurement of a speaker/driver is significantly affected by
a) the environment
b) microphone
c) amplifier (Thomann PM40c (not shown) vs Hypex UcD400MP)
d) DAC/ADC
e) process/methodology
My original post has shown that H2 is LOT lower than what a Umik-1, Sonarworks Xref20, or even Earthworks M23 can show.
Now these measurements show that level of higher order harmonics is significantly affected by the noise floor and the measurement process/method.
One interpretation is that distortion measurements published should be interpreted with a high degree of caution, due to the multiple variables affecting it. This is a kind of observer effect.
I can see why, when designing transducers, people like Lars will go to great lengths to take careful measurements, they want to measure what they are trying to measure.
Another interpretation is that measurements taken by one party cannot be directly compared with another party eg. Hobby Hi-Fi vs Klang & Ton vs Voice Coil test bench.
Finally, another interpretation is that DIYers should be cautions of own distortion measurements with if using a small 1/4" electret condenser mic.
eg. Here the W18E001 in 14L cabinet with a SB15SFCR passive radiator tuned to about 40Hz, taken with the B&K setup:
The higher order harmonics are very low andn obscured by the noise floor.
They also make the graph appear a little busy, so I do not show them for the next graphs.
Using REW's built-in EQ/DSP routines, I insert a notch filter:
Next I inserted HPF for an acoustic LR4 at 2KHz:
N.B The 3rd order peak below 2KHz remains.
Removing the resonance with a passive filter DOES REMOVE the 3rd harmonic peak caused by the resonant frequency, demonstrated by a gentleman here.
[paste]
Reference: https://www.audiosciencereview.com/...rs-for-testing-a-few-ideas.38454/post-1355264
Fantastic right? All finished?
However, his measurement ?microphone shows that H2 is higher than H3.
In my measurement with the B&K setup the H2 is so low, that H3 is still higher from 300hz to 2KHz (with or without a notch filter, active or passive)
When H3 is always higher than H2, what does that do?
Could this be the reason for how the SEAS Excel W18 (or W15/W22) sounds?
Some have described it as sounding "more detailed" .
Others have been less enamored about it, saying it is 'less musical in the long term'
Others, still have felt it had 'an inherent sound' regardless of the crossover or design eg. Thor, NaO-AEP, Orion.
eg. @wolf_teeth , @john k...
Or perhaps it's something else completely...
"The Distortion graph shows the measurement's fundamental (the linear part of its response), its harmonic distortion components up to the ninth harmonic, Total Harmonic Distortion (THD) and the level of the noise floor, which is captured before the measurement starts.
The plots are derived either from analysis of the impulse response or from stepped sine measurements.
Reference:
https://www.roomeqwizard.com/help/help_en-GB/html/graph_distortion.html#top
Lars, when say "chirp", are you referring a signal that is different signal from REW's log sine sweep?
@mark100
That is a great question and thanks for raising it. But I was was hoping someone would raise it. I'm not sure I can answer this one easily, but I was planning to summarize all my findings in my original post.
So far, going through this process clarified some things.
In the context of DIYers ie. hobbyist / non professionals- the measurement of a speaker/driver is significantly affected by
a) the environment
b) microphone
c) amplifier (Thomann PM40c (not shown) vs Hypex UcD400MP)
d) DAC/ADC
e) process/methodology
My original post has shown that H2 is LOT lower than what a Umik-1, Sonarworks Xref20, or even Earthworks M23 can show.
Now these measurements show that level of higher order harmonics is significantly affected by the noise floor and the measurement process/method.
One interpretation is that distortion measurements published should be interpreted with a high degree of caution, due to the multiple variables affecting it. This is a kind of observer effect.
I can see why, when designing transducers, people like Lars will go to great lengths to take careful measurements, they want to measure what they are trying to measure.
Another interpretation is that measurements taken by one party cannot be directly compared with another party eg. Hobby Hi-Fi vs Klang & Ton vs Voice Coil test bench.
Finally, another interpretation is that DIYers should be cautions of own distortion measurements with if using a small 1/4" electret condenser mic.
eg. Here the W18E001 in 14L cabinet with a SB15SFCR passive radiator tuned to about 40Hz, taken with the B&K setup:
The higher order harmonics are very low andn obscured by the noise floor.
They also make the graph appear a little busy, so I do not show them for the next graphs.
Using REW's built-in EQ/DSP routines, I insert a notch filter:
Next I inserted HPF for an acoustic LR4 at 2KHz:
N.B The 3rd order peak below 2KHz remains.
Removing the resonance with a passive filter DOES REMOVE the 3rd harmonic peak caused by the resonant frequency, demonstrated by a gentleman here.
[paste]
Reference: https://www.audiosciencereview.com/...rs-for-testing-a-few-ideas.38454/post-1355264
Fantastic right? All finished?
However, his measurement ?microphone shows that H2 is higher than H3.
In my measurement with the B&K setup the H2 is so low, that H3 is still higher from 300hz to 2KHz (with or without a notch filter, active or passive)
When H3 is always higher than H2, what does that do?
Could this be the reason for how the SEAS Excel W18 (or W15/W22) sounds?
Some have described it as sounding "more detailed" .
Others have been less enamored about it, saying it is 'less musical in the long term'
Others, still have felt it had 'an inherent sound' regardless of the crossover or design eg. Thor, NaO-AEP, Orion.
eg. @wolf_teeth , @john k...
Or perhaps it's something else completely...
Not sure what you are calling my attention for, but 3rd order is not a pleasing even order distortion where extra warmth, detail, or air is presented. Odd order HD just sounds wrong, not extra detail.
My additional comment to this 3rd order higher prevalence, is that if -50dB is the perceived threshold of inaudibility, that it won't matter if either is higher magnitude over the other if the are both below that output threshold.
My additional comment to this 3rd order higher prevalence, is that if -50dB is the perceived threshold of inaudibility, that it won't matter if either is higher magnitude over the other if the are both below that output threshold.
Ben,
I thought I would ask for your input because you have had a wealth of listening experiences with all manner of speakers, big and small, and a large variety of drivers.
Would you say that the SEAS Excel magnesium cone drivers have a distinct flavour, or sound?
What about after the resonance was eliminated?
And how does this compare to the SEAS Nextel cone drivers, which share the same motor (with a copper ring above, and below the pole piece, but not a copper sleeve through it (eg. original Scan-Speak SD, Purifi motors)
I’ve come across this threshold of audibility being -50dB before. Where does it come from again? I’d like to learn more.
Or, is perhaps the inability to measure below -50dB leading to the interpretation that anything below doesn’t seem to matter as much?
because the microphone all have difficulty measuring below -50dB, that seems to be the threshold.
The Omni1 and MD42 is the first microphone I’ve encountered that can characterise ~-60dB, the others are affected by self noise or capsule distortion or both.
It would seem to me, that one needs a measurement chain (from DAC + amp to device to mic to ADC) that has, -x dB of harmonic and intermodulation and other kinds of non-linearity.
And then, via software, inject various forms of non-linearity into it.
One test via headphones, another whilst in an anechoic room eg. with a noise floor of 10dB(A), and then in your room listening room. With a test tone, or even better, with your own music.
I wonder if this is already being done in automotive space, that is where all the R&D seems to be… the best sound systems will be in the stationary EV.
I thought I would ask for your input because you have had a wealth of listening experiences with all manner of speakers, big and small, and a large variety of drivers.
Would you say that the SEAS Excel magnesium cone drivers have a distinct flavour, or sound?
What about after the resonance was eliminated?
And how does this compare to the SEAS Nextel cone drivers, which share the same motor (with a copper ring above, and below the pole piece, but not a copper sleeve through it (eg. original Scan-Speak SD, Purifi motors)
I’ve come across this threshold of audibility being -50dB before. Where does it come from again? I’d like to learn more.
Or, is perhaps the inability to measure below -50dB leading to the interpretation that anything below doesn’t seem to matter as much?
because the microphone all have difficulty measuring below -50dB, that seems to be the threshold.
The Omni1 and MD42 is the first microphone I’ve encountered that can characterise ~-60dB, the others are affected by self noise or capsule distortion or both.
It would seem to me, that one needs a measurement chain (from DAC + amp to device to mic to ADC) that has, -x dB of harmonic and intermodulation and other kinds of non-linearity.
And then, via software, inject various forms of non-linearity into it.
One test via headphones, another whilst in an anechoic room eg. with a noise floor of 10dB(A), and then in your room listening room. With a test tone, or even better, with your own music.
I wonder if this is already being done in automotive space, that is where all the R&D seems to be… the best sound systems will be in the stationary EV.
Last edited:
I've heard several of the AlMg Excel lineup drivers at DIY events, but not a Nextel paper unless in a commercial design. The Excels I've heard are very clean if done right, and the metal cone breakup is audible if not suppressed, especially on the W22. They will ring like a bell as most metal cones will.
The threshold of audibility has been stated many times and I've just held onto the information. -25dB is the minimum suppression of problematic artifacts for most people. -40dB is 1% HD, and typically highly preferred over -25dB. -50dB is 0.5% HD and typically deemed inaudible in reference to the nominal signal. It is rated by what we can perceive, not what we can measure.
The threshold of audibility has been stated many times and I've just held onto the information. -25dB is the minimum suppression of problematic artifacts for most people. -40dB is 1% HD, and typically highly preferred over -25dB. -50dB is 0.5% HD and typically deemed inaudible in reference to the nominal signal. It is rated by what we can perceive, not what we can measure.
I would tend to agree with that observation.
The REW exponentially swept sine measurement does seem to be excessively smoothed out at high frequencies, especially when compared with the high-resolution STEPS stepped-sine measurement with 48 points per octave. The REW measurement looks very unusual if compared to measurements of similar drivers obtained using MLS-based measurements using MLSSA or IMP/M measurement systems. I wonder if someone can explain this behaviour.
The STEPS stepped-sine 48-points per octave measurement also seems to have some unusual behaviour. Looking at the frequency response in the
2kHz to 4kHz frequency band, there are quite noticeable peaks and troughs in the measurement. The variation is about 2dB, which seems to be very large. The frequency spacing is also quite narrow. Are there really that many diaphragm breakup modes in that frequency range?
Assuming that the DAC being used as part of the measurement system has an appropriately configured anti-aliasing filter, how can stepped-sine samples at discrete frequencies within the passband give aliasing? I'm not sure what is being referred to in this instance.
Doesn't a stepped-sine transfer function analyzer make use of bandpass filters around the measured response? At least I think the ones from the 1980s did. That is one of its benefits, and it allows the method to have a high signal-to-noise, even under potentially noisy test conditions (e.g., ground vibration testing of an aircraft in a hangar). This type of analyzer can also measure phase response differences between input and out, typically with better than 1° accuracy.
Last edited:
Yes, we assume that the ADC and DAC in the chain avoids the aliasing from sampling in the time domain.
I was referring to the sampling in the frequency domain that the stepped sine method performs. The unknown frequency response of the DUT is being sampled and the sampling theorem applies here as well.
You did measurements at 1cm and 40cm which quite a big step of 32 dB.
Would it make sense to make finer steps to find a sweet spot where H2 and H3 measurement would be optimal?
Would it make sense to make finer steps to find a sweet spot where H2 and H3 measurement would be optimal?