soongsc said:
No, no, no. For well designed class A or AB with error correction they DO NOT INCREASE!! Only poor designs do. Unfortunately, this is the case of most designs, that they ARE poor.
Hi
Well I am happy having discovered this PSU issue actually. I am working on a gyrator type voltage stabilisation for the first stages to overcome it permanently. It will be interesting to see if there is any difference in the sonic pattern happening.
One thing I love of that MOSFET's compared to the NAIM's is the complete lack of harshness. This is not to say that the NAIM's sound harsh but if you are very familiar with it you know there is something "behind the curtain".
I discovered this many years ago and its the first time I saw this proofed to some extent in my measurements with slightly present 2nd harmonics.
Sadly I don't have someone that can borrow me a NAP 250 as this is basically the same as the NAP 110 / 140 design but has fully regulated power supply.
I am curious.
With the measurements I did there is kind of " distinguishing filter " between arbitrary noise and recurring signals due to averaging .
Though not being a measurement whiz, I can't really imagine any other technique to sort out signals from the noise floor.
Even if you set the rectangular window EXACTLY to fit the stimulus with your special technique, there should remain artefacts of any other signals present that are NOT exact multiples of the stimulus ( like the hum seen ), no?
Maybe I didn't understand your approach very well.
Greetings
Michael
gedlee said:
Well I am happy having discovered this PSU issue actually. I am working on a gyrator type voltage stabilisation for the first stages to overcome it permanently. It will be interesting to see if there is any difference in the sonic pattern happening.
One thing I love of that MOSFET's compared to the NAIM's is the complete lack of harshness. This is not to say that the NAIM's sound harsh but if you are very familiar with it you know there is something "behind the curtain".
I discovered this many years ago and its the first time I saw this proofed to some extent in my measurements with slightly present 2nd harmonics.
Sadly I don't have someone that can borrow me a NAP 250 as this is basically the same as the NAP 110 / 140 design but has fully regulated power supply.
I am curious.
With the measurements I did there is kind of " distinguishing filter " between arbitrary noise and recurring signals due to averaging .
Though not being a measurement whiz, I can't really imagine any other technique to sort out signals from the noise floor.
Even if you set the rectangular window EXACTLY to fit the stimulus with your special technique, there should remain artefacts of any other signals present that are NOT exact multiples of the stimulus ( like the hum seen ), no?
Maybe I didn't understand your approach very well.
Greetings
Michael
mige0 said:
Actually, this is not correct. The noise is random, will average to zero over a long enough time. The hum is not coherent with the window, and IF the signal is, then the hum will also average down to zero.
In your example, the hum was very nearly coherent with the signal. You could not have averaged out the hum without averaging out the signal.
Earl, please check your mailbox (I mailed you at the given address).
@all: I successfully undertook a measurement according to the principles given by Dr.Geddes. Noise floor is 20dB lower than it ever was, for the first time I can see the harmonics of a -90dBfs signal with my 24bit soundcard in loopback....
- Klaus
@all: I successfully undertook a measurement according to the principles given by Dr.Geddes. Noise floor is 20dB lower than it ever was, for the first time I can see the harmonics of a -90dBfs signal with my 24bit soundcard in loopback....
- Klaus
Hi
Klaus, can you show us some plots ?
At what absolute level ( dBV ) did you take the measurements.
Greetings
Michael
KSTR said:
Klaus, can you show us some plots ?
At what absolute level ( dBV ) did you take the measurements.
Greetings
Michael
Hi,
I tried this approach (or at least what I think it is), using the hints given by Mr.Geddes and writing a C quickie according to that.
Now what can I say.... THIS IS INCREDIBLE!
Please look at the attached plots, showing loopback measurements of my soundcard, at 24bits and 48kHz. The top plot is trying to FFT (16K block size, no windowing) a -90dBfs-Signal in the conventional way, the center plot is what I got using the Geddes method, and the bottom plot is the stimulus alone, for reference. The sound card has 0dBfs == +6dBV levels. It uses an AK4524 24/96 2ch-codec which has analog input gain adjustment, while output gain is controlled digitally. Therefore I leave the output gain always at 0dB, but the input gain can be adjusted without penalty over a +-18dB range, both noise floor and signal shift up and down in unison. The outputs are buffered/filtered with MC33078 op-amps. The inputs are AC-coupled with a cheap 10uF electrolytic, AC-input-impedance is 10k (DC: 15k) and there is a resistor divider to accomodate the +6dBV input range (because +-2.83Vpeak would exceed the range of the codec, running from +5V).
With the conventional method, even with massive averaging, the noise floor is a tad above -140dBfs and there is no way to indentify any harmonics there. We can see some spuriae at ~300Hz, 3.4kHz, and the 50Hz hum. With very much guesstimating one might identify a tiny wiggle at 2kHz, which could be the 2nd harmonic... or just anything else.
With the Geddes method (averaging some 1800 blocks, which would in theory lower random noise by 32dB if I get that right?), I could lower the average noise floor to -165dBfs, this is more than 20dB better (the claimed increase in resolution) and only some 25dB above the inherent test signal limit of -180dBfs, using 24 bit resolution. The hum components and other unsynced spuriae are gone, as one would expect. 2nd and 3rd harmonic are cleary indentified as being about 50dB down from the fundamental, also H7, H9, H11, H13 are visible. Indicated THD was a reasonable 0.45%, while THD with full-scale signals is around 0.0004%, the harmonics more than 110dB down.
Note that the test frequency was ~987Hz, not 1kHz, therefore the harmonics shift a little to the left, compared to the 1kHz grid. The peak at precisly 12kHz is fs/3 (fs=48kHz). Some other non-harmonic artifacts can be seen, these might be alias products and/or IM components with the sample frequency (the stimulus frequency in this quick test wasn't optimized in this regard. But I used a prime multiple of the FFT resolution frequency to get the generator and converter distortion down as low as possible). Also note that I upped the input level a little to get close to -90dBfs reported level, but that was uncalled for because the stimulus was generated to have -90dBfs and then I also got -95dB reported from the analyser when I FFT'd the stimulus, after the measurement. That "error" is an effect of the chosen FFT parameters, but I wasn't aware of that. Measuring took quite a while, I recorded the signal for 10 minutes, and the program worked another 4 minutes on the data.
This really is a great improvement in resolution, and this was only my first try, nothing optimized. Such a simple idea, but truly ingenious. Thank you, Mr.Geddes!
Regards, Klaus
I tried this approach (or at least what I think it is), using the hints given by Mr.Geddes and writing a C quickie according to that.
Now what can I say.... THIS IS INCREDIBLE!
Please look at the attached plots, showing loopback measurements of my soundcard, at 24bits and 48kHz. The top plot is trying to FFT (16K block size, no windowing) a -90dBfs-Signal in the conventional way, the center plot is what I got using the Geddes method, and the bottom plot is the stimulus alone, for reference. The sound card has 0dBfs == +6dBV levels. It uses an AK4524 24/96 2ch-codec which has analog input gain adjustment, while output gain is controlled digitally. Therefore I leave the output gain always at 0dB, but the input gain can be adjusted without penalty over a +-18dB range, both noise floor and signal shift up and down in unison. The outputs are buffered/filtered with MC33078 op-amps. The inputs are AC-coupled with a cheap 10uF electrolytic, AC-input-impedance is 10k (DC: 15k) and there is a resistor divider to accomodate the +6dBV input range (because +-2.83Vpeak would exceed the range of the codec, running from +5V).
With the conventional method, even with massive averaging, the noise floor is a tad above -140dBfs and there is no way to indentify any harmonics there. We can see some spuriae at ~300Hz, 3.4kHz, and the 50Hz hum. With very much guesstimating one might identify a tiny wiggle at 2kHz, which could be the 2nd harmonic... or just anything else.
With the Geddes method (averaging some 1800 blocks, which would in theory lower random noise by 32dB if I get that right?), I could lower the average noise floor to -165dBfs, this is more than 20dB better (the claimed increase in resolution) and only some 25dB above the inherent test signal limit of -180dBfs, using 24 bit resolution. The hum components and other unsynced spuriae are gone, as one would expect. 2nd and 3rd harmonic are cleary indentified as being about 50dB down from the fundamental, also H7, H9, H11, H13 are visible. Indicated THD was a reasonable 0.45%, while THD with full-scale signals is around 0.0004%, the harmonics more than 110dB down.
Note that the test frequency was ~987Hz, not 1kHz, therefore the harmonics shift a little to the left, compared to the 1kHz grid. The peak at precisly 12kHz is fs/3 (fs=48kHz). Some other non-harmonic artifacts can be seen, these might be alias products and/or IM components with the sample frequency (the stimulus frequency in this quick test wasn't optimized in this regard. But I used a prime multiple of the FFT resolution frequency to get the generator and converter distortion down as low as possible). Also note that I upped the input level a little to get close to -90dBfs reported level, but that was uncalled for because the stimulus was generated to have -90dBfs and then I also got -95dB reported from the analyser when I FFT'd the stimulus, after the measurement. That "error" is an effect of the chosen FFT parameters, but I wasn't aware of that. Measuring took quite a while, I recorded the signal for 10 minutes, and the program worked another 4 minutes on the data.
This really is a great improvement in resolution, and this was only my first try, nothing optimized. Such a simple idea, but truly ingenious. Thank you, Mr.Geddes!
Regards, Klaus
Attachments
KSTR said:
You show very well what the technique is capable of. It allows you to see what cannot be seen otherwise.
Good job!
Hi
Klaus, this looks great indeed. Do you think I could perform this as well though I'm NOT firm in C++ ? What do I need ?
I tried to teak my setup.
I can come somewhere close (10 dB) to your limits when I lower the sensitivity of my line-IN to around 0 dBFS = 20 dBV AND use max resolution.
With 262 144 FFT points the noise floor is down at roughly –155 dBFS. The measurement took less than a minute with averaging set to 5.
Measurement is in loop back mode . Y-axis in dBFS is valid for ALL plots. Input frequency was shifted for clarity.
Concluding from the different settings shown in the plot I would expect a ten times crancing the resolution up would lower the noise floor another 10 dB.
I'll might ask if 2x10^6 FFT points could be implemented though right now I can't think of any real benefit from that.
An other way to look deep into the noise floor is shown in post 178 by engaging a good ( mic- ) pre amp.
http://members.aon.at/kinotechnik/diyaudio/dipol/space/geddes/mackie_ONYX_400F_MIC_40dB.gif
Note that Y-axis is scaled in dbV there.
Greetings
Michael
KSTR said:
Klaus, this looks great indeed. Do you think I could perform this as well though I'm NOT firm in C++ ? What do I need ?
I tried to teak my setup.
I can come somewhere close (10 dB) to your limits when I lower the sensitivity of my line-IN to around 0 dBFS = 20 dBV AND use max resolution.
With 262 144 FFT points the noise floor is down at roughly –155 dBFS. The measurement took less than a minute with averaging set to 5.
An externally hosted image should be here but it was not working when we last tested it.
Measurement is in loop back mode . Y-axis in dBFS is valid for ALL plots. Input frequency was shifted for clarity.
Concluding from the different settings shown in the plot I would expect a ten times crancing the resolution up would lower the noise floor another 10 dB.
I'll might ask if 2x10^6 FFT points could be implemented though right now I can't think of any real benefit from that.
An other way to look deep into the noise floor is shown in post 178 by engaging a good ( mic- ) pre amp.
http://members.aon.at/kinotechnik/diyaudio/dipol/space/geddes/mackie_ONYX_400F_MIC_40dB.gif
Note that Y-axis is scaled in dbV there.
Greetings
Michael
Science of hearing distortions
I know that this is an old thread, but anyway it seems that my question belongs here.
Is there some (scientific) evidence on how small crossover distortion human is capable to hear (a) consciously or (b) unconsciously? In my question "hearing" means that signal (sound) reaches ear and then nerves bring it to brains so that brains do at least something with it.
I know that this is an old thread, but anyway it seems that my question belongs here.
Is there some (scientific) evidence on how small crossover distortion human is capable to hear (a) consciously or (b) unconsciously? In my question "hearing" means that signal (sound) reaches ear and then nerves bring it to brains so that brains do at least something with it.
The problem is the question is ill-posed. "crossover distortion" can take many forms and each form would have different numbers. In general thios form of nonlinearity is highly audible because its actually % value rises as the signal level falls. This means that it become more and more audible at ever lower signal levels until the signal falls into the noise floor (and even then its likely to be audible). If you read my papers on nonlinear distortion perception you will find a metric for evaluating this situation that should work fairly well. Thats as scientific as I know.
How small is small enough?
I will state the question other way: How small crossover distortion is small enough?
I will state the question other way: How small crossover distortion is small enough?
How does this 'because' follow from what precedes it? The percentage can go up as the signal level decreases whilst the absolute level of the distortion components remain the same.
Since 'noise floor' is bandwidth dependent, there's no such thing as 'falling into the noise floor', its just a convenient red herring. On an FFT analysis, the noise floor will depend on the width of each measurement bin - more bins, lower apparent noise floor but no change in the noise itself.
Are they to be found on your website?
Thats not much better. In my papers I show a sclae of when subjects will just "detect" this distortion, find it "clearly audible" and finally find it "objectionable". Is "detect", "good enough"? How about "clearly audible", but not necessarily "objectionable"? It's a continuous scale and you have to test the system using the appropriate metric and compare that number to the scale. Otherwise you are just guessing, and since you are asking the question I suspect that you don't want to guess.
And I sincerely hope that you are not looking for a "THD" number because if thats the case, there isn't one.
And I sincerely hope that you are not looking for a "THD" number because if thats the case, there isn't one.
None of this "follows" if you have not read and understood the papers on my web site.
Your "noise floor" argument is incorrect. If the noise is external from the FFT calculation then it is unaffected by the bin width. Only in the case of internal noise from the FFT calculations (bit depth, sample stability, etc.) would the noise be a function of the bin width. The noise floor in an external device has a level that is independent of how you measure it, so the "how" cannot change the measured level or you are doing something wrong.
Well of course its obvious that the noise is unaffected by some change to the mathematical algorithm used in its measurement. But that is not what I said. My meaning was that if the mag/phase in the FFT bins is plotted out in modulus form then those moduli will be bin width dependent. How much noise gets in to each one depends on how wide it is.
No, the noise in an external resistor has a level dependent on the measurement bandwidth. SQRT(4kTRB) remember? Unless we know B we can't get the value. The noise floor in an FFT plot therefore does vary with the number of points in the FFT used to create it because more bins (assuming same sample rate) means reduced B.
Avoiding my first question because you have no reasoned response?😀
I have to differ slightly on the behavior of crossover distortion at low levels. There is a common misconception that crossover distortion percentage continues to increase as level decreases. This is not generally true for a properly designed and biased class AB amplifier. There is a power level at which the crossover distortion peaks, and then decreaes as level is decreased. This power level depends on the bias current in the class AB amplifier. At some point, as power decreases, the class AB amplifier enters its class A region of operation, where crossover distortion will have decreased and eventually disappeared for all practical purposes. This often occurs at a fairly small power level, but not a miniscule one.
Consider an optimally-biased class AB amplifier rated at 150W and using two pairs of output transistors, all with 0.22 ohm emitter resistors. Total bias current will be about 240 mA. This means that the amplifier will be able to put out about 480 mA into the 8 ohms load before it exits its class A region. This corresponds to about 1 watt into 8 -ohms. Small, but certainly not miniscule, and certainly not below any reasonable SNR floor.
Most people focus on static crossover distortion, as above.
There is also dynamic crossover distortion, which can be more insidious. This has to do with the ability to turn output transistors off fast enough for the needs of the signal. It often increases with operating level. It is largely a high-frequency phenomena. It is also frequently called switching distortion. So-called "non-switching" amplifiers DO NOT necessarily avoid this kind of distortion. Avoiding dynamic crossover distortion requires fast output transistors and usually high driver transistor bias currents to enable good suck-out of stored charge in the base of the output transistor. The so-called "speed-up" calacitor sometimes used between the bases of the output transistors is an attempt to reduce dynamic crossover distotion. If you have an amplifier whose output stage current draw increases when you go up to full output at 20 kHz, this is a sign that there is likely dynamic crossover distortion occurring, as it is associated with totem-pole conduction where the opposite transistor turns on before the first transistor has been fully shut off.
Although not completely immune to it, well-designed MOSFET power amplifiers suffer less from dynamic crossover distortion because the MOSFETs are fast and because they do not involve minority carrier charge storage.
Cheers,
Bob
You are talking apples versus oranges and arguing about something which you do not seem to understand. The equation that you give for thermal noise is for the rms value of measured over a bandwidth B. B IS NOT the bandwidth of the bins (this is an RMS measure) but the bandwidth of the total measurement independent of the number of bins. Its an rms value NOT a spectrum value. You need to understand the difference if you want to continue this discussion.
Let's say that the noise level in an amp is -30dB SNR (its a bad amp) and lets assume that this is not thermal noise so your example does not apply. If I measure this noise level with an FFT (or anything else) then I'd better get -30dB SNR regardless of the FFT size, or bandwidth or anything else. It's -30 dB SNR regardless of how I measure it. Now the actual noise RMS level will increase as the bandwidth increases, of course, (but so does the signal) but the level in any particular bin DOES NOT increase. RMS is bandwidth dependent, but the sprectum level is not.
I'm not going to argue with you about a paper that you haven't read.
Let's say that the noise level in an amp is -30dB SNR (its a bad amp) and lets assume that this is not thermal noise so your example does not apply. If I measure this noise level with an FFT (or anything else) then I'd better get -30dB SNR regardless of the FFT size, or bandwidth or anything else. It's -30 dB SNR regardless of how I measure it. Now the actual noise RMS level will increase as the bandwidth increases, of course, (but so does the signal) but the level in any particular bin DOES NOT increase. RMS is bandwidth dependent, but the sprectum level is not.
I'm not going to argue with you about a paper that you haven't read.
Bob, the difference is, as you say, the fact that I was refering to an idealized case of static distortion and what would happen. In a real situation things are more complex as you suggest.
The point is that in considering audibility of any type of distortion one has to consider the masking effect of the signal. Crossover distortion is highly audible because there is less signal to mask it (or its not "crossover distortion"). This is true no matter how the actual distortion levels behave in specific.
So you say. Easy to say, hard to show.
It could be either the bandwidth of one single bin or it could be the total measurement bandwidth. But that was my point - without a defined bandwidth then your phrase 'noise floor' is meaningless.
I think what you must be meaning here is you want me to accept your word for something if I want you to continue in the discussion. I really don't care if you continue or not though, so this is irrelevant.
But how do we know its SNR is -30dB before we measure it? You seem to have put the cart before the horse here - the SNR must be unknown prior to measurement or why on earth would anyone waste time with making a measurement? But if you want to say 'here's one I measured earlier, its -30dB' then I'll ask 'what bandwidth did you use?'.
So it has no thermal noise whatsoever - the measurement bandwidth has no impact at all? Doesn't sound like a real world thought experiment at all.
Ah, so it does indeed have some thermal noise then. At least we've got this clear.
I agree subject to the proviso that the bandwidth of each bin remains constant.
But this is yet another red herring. I'm arguing about what you've posted here. And you've still not answered my first question, just brushed it aside.
- Status
- Not open for further replies.