# Understanding Audio Measurements

Actually, at low levels, the analog nonlinearities are minimized. They're most important at high levels.

I don't disagree in principle, but it would be interesting to perform a sanity check every now and then.

I think the assumption is made that at -20dB the DAC performs in an ideal manner and the output voltage is recorded, the rest of the level measurements are normalized based on that point.

Normalised how? 0dB f first step equals -20db, -10dB of 2nd step equals -30dB and so on? Or.. ?

Normalised how? 0dB f first step equals -20db, -10dB of 2nd step equals -30dB and so on? Or.. ?

Linearity deviation is assumed to be 0dB at -20dB output, you record the output voltage at that level, then you calculate expected voltages at other levels and compare them to actual and plot deviations on the linearity graph, something like that. Or I may all be wrong

Well it is by most definitions the start of the midrange. But I'm not sure what @Soniclife defined as midrange in his measurements.

It is worth noting though that bass noise floor will tend to mask the midrange too, although not to the extent it masks bass.
I don't think I'll get involved in 'where does the midrange start?", last time I looked it up there was no consensus .
My noise floor measurment is at https://audiosciencereview.com/foru...140-the-limit-in-measurements.2700/post-76409, seems like my room goes negative around 300Hz, so well below the frequencies where the ear is most sensitive to avoid masking.

Ok, that sounds reasonable. But noise floor at -120dB seems one thing and amplitude error of +/- 0.5dB of a signal at -120dB seems another, and that last one doesn't seem like a hearable issue at all to me. Btw, wouldn't even a very solid amp have a SNR of 100-110 dB so its noise will anyhow mask those little errors?
Essentially all noise is filtered with linearity tests with very narrowband filters. Without that, you would be right.

Linearity deviation is assumed to be 0dB at -20dB output, you record the output voltage at that level, then you calculate expected voltages at other levels and compare them to actual and plot deviations on the linearity graph, something like that. Or I may all be wrong
You are correct.

You are correct.

Well, that is the same what I meant: 2V of max output corresponds to 0dB. Based on that you take voltage of -20dB as a new 0dB level and from there you go down in -10dB increments, right? That makes -100dB level of your linearity measurement actually pretty close to the noise floor, doesn't it?

Essentially all noise is filtered with linearity tests with very narrowband filters. Without that, you would be right.

But noise is broadband, so noise (and other garbage) caught between 200Hz filter limits is still there, isn't it?

But noise is broadband, so noise (and other garbage) caught between 200Hz filter limits is still there, isn't it?
I only allow 20 Hz worth of bandwidth for the filter. So take your total spectrum of noise and divide it down by 20 Hz and it becomes a very small number. I have done it with 1 Hz too and the results are the same.

I only allow 20 Hz worth of bandwidth for the filter. So take your total spectrum of noise and divide it down by 20 Hz and it becomes a very small number. I have done it with 1 Hz too and the results are the same.

Sure. But if the noise+garbage is buzzing at all frequencies at say -120dB then there is some noise present in the range of 190-210Hz also at the -120dB level, right?

That would also explain why it it stays the same at 1kHz as well., correct?

And the fact that D50 has better linearity than D10? Well, it uses better DAC and better op amps so it's noise floor is lower.
The fact that your measurements fluctuate each time you make a new measurement? That could be because of the random nature of the noise.

Or..?

If that is the case than we should really be ignoring those irregulatiries below -100dB (taken at -20dB signal reference) as they are showing no linearity problem exists as long as you don't hit the noise floor. I believe somebody from the DAC world already said that here, isn't that so?

Well, that is the same what I meant: 2V of max output corresponds to 0dB. Based on that you take voltage of -20dB as a new 0dB level and from there you go down in -10dB increments, right? That makes -100dB level of your linearity measurement actually pretty close to the noise floor, doesn't it?

You misunderstood, he measures the voltage at -20dB, say it is 199mV, this is assumed to be 0dB *deviation* from linearity, under this assumption the expected voltage at 0dB *output* is 1990mV, 1990mV is then compared to the actual voltage at 0dB output which gives a point for the 0dB level on the linearity plot, and so on for all levels.

You misunderstood, he measures the voltage at -20dB, say it is 199mV, this is assumed to be 0dB *deviation* from linearity, under this assumption the expected voltage at 0dB *output* is 1990mV, 1990mV is then compared to the actual voltage at 0dB output which gives a point for the 0dB level on the linearity plot, and so on for all levels.

Ok, and how does he measure the level at -10dB, -20dB and so forth..? Does he measure the voltage at -30dB and assume that to be the level at -10dB?

Last edited:
Ok, and how does he measure the level at -10dB, -20dB and so forth..? Does he measure the voltage at -30dB and assume that to be the level at -10dB?

-20dB is the first point, it's always 0dB deviation, note it is 0dB *deviation*, not level. Knowing the voltage at -20dB it's easy to arrive at expected voltages at all possible levels using log arithmetic. He then measures the voltages at -10, -20 and so forth and calculates the deviation from those calculated values. You must establish a reference point first. You can do it at 0dB too but chances are -20dB is in a more linear area of the device so you get a more correct picture of +/- deviation overall. Maybe someone has a better way to explain this.

-20dB is the first point, it's always 0dB deviation, note it is 0dB *deviation*, not level. Knowing the voltage at -20dB it's easy to arrive at expected voltages at all possible levels using log arithmetic. He then measures the voltages at -10, -20 and so forth and calculates the deviation from those calculated values. You must establish a reference point first. You can do it at 0dB too but chances are -20dB is in a more linear area of the device so you get a more correct picture of +/- deviation overall. Maybe someone has a better way to explain this.

Ok, got it. So, according to this, when does the measured THD+noise floor level start to interfere with this measurement?

Ok, got it. So, according to this, when does the measured THD+noise floor level start to interfere with this measurement?

When the noise floor in the band being measured gets within about 25 db of the signal level. With a 20 hz wide filter Amir mentions that means with a device having around a 105 db SNR, you will start seeing some effect at about the -100 dbFS point of the linearity test.

Here's a demonstration of the effect of noise with the filter Amir uses.

I have no real desire to chase down the rabbit hole of this thread but have a couple of comments on in-room noise:
1. Our ear/brain averages signals and organic processing allows us to hear correlated signals below the noise floor. In some cases IIRC our brain will help "fill-in" signals that we anticipate should be there.
2. In-room noise comes in a variety of flavors. Some is correlated, some is not, and some is a mixture of both. Examples:
• Transformer buzz happens at 60 or 120 Hz (in the USA; may be 50 or 100 Hz in your neck of the woods);
• SMPS are wideband noise sources, usually well above the audio band, but subharmonics can creep in and of course EFI/RFI can be modulated to audio by other components in the system, and those often appear as signals with a little spreading around whatever center frequency they happen to hit (note LED-based replacements usually have SMPS inside);
• Fluorescent lights with those little fast-start ballasts operate up around 60 kHz (tore some hair out chasing that down years ago) and again can be modulated to baseband;
• Dimmers tend to be broadband RF noise sources but often exhibit fundamentals and subharmonics in the audio band;
• Motors like HVAC fans tend to have all sorts of signals, from hum at 60 Hz or so to tones around the rotational frequency and wind noise modulated and influenced by the ducts (which can act like resonant pipes -- think organ pipes), and may appear as a combination of narrow- and broad-band noise sources in the room;
• Etc.
Bottom line is that some noises we can hear right through; others inject specific tones or bands of noise that may only impact certain notes during playback. We may have that 116 dB or so of dynamic range in a quiet room for some notes, not for others, even if the noise on a broadband SPL meter is in the 20 - 40 dB range.

IME/IMO - Don

When the noise floor in the band being measured gets within about 25 db of the signal level. With a 20 hz wide filter Amir mentions that means with a device having around a 105 db SNR, you will start seeing some effect at about the -100 dbFS point of the linearity test.

And that is exactly what seems to be happening. That would make noise the reason for non linearity at low dbFS and it makes a perfet sense to me.

Replies
108
Views
12K
Replies
7
Views
702
Replies
2
Views
2K
Replies
57
Views
15K
Replies
49
Views
19K