RE Victor's amazing oscillator- It represents a lot of detailed component tuning and selection. My first gen example get -150 dB H2 and -155 dB H3. The current examples are even better. Measuring that is not easy. IVX's instincts that the impedance needs to be lowered to lower the noise is on track but there are other noise sources you need to consider. 600 Ohms limits the noise to around 3 nV or -170 dBV/1 Hz band. Most analyzers I have checked have an equivalent input noise of around 5 nV/rtHz (Approx 1K Ohms in series with the source). For THD+N that sets limits. Its around -130 dBV with a 20 KHz limit and -125 dBV with the more common 30 KHz limit.
I have seen the improved results with a 10V source. I modded my KH4400 to run at 10V RMS and it does help along with the 50 Ohm source.
Using analog frequency domain filters is the only way to get really low distortion. My CLT-1 gets -170 dB HD3 with heroic passive LC filters. Its a solution that will work for a DAC (the internal oscillator of the CLT-1 starts at .1% (-60 dB) and uses filters to get to -170). Maybe posting a simple to build circuit to filter a "generic" DAC to -150 dB distortion would be useful?
I have seen the improved results with a 10V source. I modded my KH4400 to run at 10V RMS and it does help along with the 50 Ohm source.
Using analog frequency domain filters is the only way to get really low distortion. My CLT-1 gets -170 dB HD3 with heroic passive LC filters. Its a solution that will work for a DAC (the internal oscillator of the CLT-1 starts at .1% (-60 dB) and uses filters to get to -170). Maybe posting a simple to build circuit to filter a "generic" DAC to -150 dB distortion would be useful?
CG, I agree about the standalone LPF > 2ord because a decent DAC pretty much everyone already has. Regarding AP analyzers I think it is far not the top at all, and I never had the goal to repeat the performance of the best AP, only outperform that.
PS: regarding the "Ferrari" I think it rather a gold plated iPhone 2007 ))
PS: regarding the "Ferrari" I think it rather a gold plated iPhone 2007 ))
Last edited:
100% correct regarding Victor's osc but I don't know such AP which guarantees even -140db H2 or H3.
On the other hand, the LPF with residual H3 -170db(-155db -15db=-170db) it becomes a challenging task as well.
The LPF with H3 -150db is relatively easy to get, so I prefer to wipe DAC's -135db down to -150db with a simple sallen-key LPF
and next add anti-H2 and anti-H3 harmonics digitally. I have some ideas about how to do that in one click i.e. automatically but not ready
to show results yet.
PS: the wien-bridge oscillator anyway keeps some disadvantages vs DAC+LPF:
a) a huge jitter skirt, the main reason why most of FFT software wrongly calculates THD+N with such sine sources(if there isn't implemented AES17 notch)
b) it impossible to use in the automatic level sweep tests
c) each unit requires a recursive adjustment F0 and even swap the opamp to get better H2 and so on
Last edited:
In the last months, I've worked on a bibliography covering a wide range of electrical oscillator: bibliography_electrical_oscillators.pdf
While mostly concerned with RF oscillators, the document also includes a list of references discussing low distortion RC oscillators.
While mostly concerned with RF oscillators, the document also includes a list of references discussing low distortion RC oscillators.
The APU is great tool, but "inexpensive" could mean not the same for somebody - here in EU anounced 100$/70$ for APU boxed/PCBA is around 170EUR/97EUR + 13EUR what we have to pay in Ali shop (and wait few weeks)...
So, I understand that selling such a special tool is not in a large amounts, even if the price is reflexing the effort + labor + material you have put into it.
Good luck with a Cosmos DAC.
Martin
A worthy effort. Thank you.
FYI: first page of references, "[Sch55] Alexander Schure. Crystall Oscillators", should maybe be spelled with one "L"? (Two Ls is not wrong but is extremely "obsolete" in modern English; 'For the most part it resembleth Crystall: which is the reason that some hath tearmed it the root of Crystall. William Gilbert of Colchester', late 1600s) Yes, I learned OCD spelling at my mother's knee.
Also, this may just be Adobe PDF being stupid, but in extracting text there are mis-encodings or something. These do not affect eye-reading but may foil mechanical indexing.
[/I]Very nice list, thanks a lot. Did you consider including the well known Mr. Janasek's paper http://www.janascard.cz/PDF/An ultra low distortion oscillator with THD below -140 dB.pdf ? I find it quite informative.
zergxia, actually John did correct his REW in May 2021, so 2022 versions should be Ok. However, he extremely simplifies so-called "AES17 notch" needed for THD+N, REW simply ignores H1+/-40% but AP2700 for instance, does that like that:
So, even REW, and Arta aren't perfectly matched vs AP regarding the THD+N. Other FFT apps seem never heard about AES17_notch requirements at all, hence, they deserve to be ignored, in my opinion ))
So, even REW, and Arta aren't perfectly matched vs AP regarding the THD+N. Other FFT apps seem never heard about AES17_notch requirements at all, hence, they deserve to be ignored, in my opinion ))
I looked up the definition of AES-17 notch filter, attached below. Its pretty broad with a Q of 1.2 to 3. This is in the digital audio spec which may be why it's not well known. Its actually modeled on the AP and the Shibasoku analog notch filters. I worked with Matt at Quantasylum on this exact issue. Digital sources are pretty good on close in phase noise compared to analog RC oscillators. Its much easier I gather to just blank areas around the fundamental and depending on the selected window you won't get any leakage from the fundamental.
The band limit is specified to be 20 KHz with a really steep rolloff, 0 dB @ 20 KHz, 60 dB @ 24KHz. Not remotely easy in the analog domain.
The band limit is specified to be 20 KHz with a really steep rolloff, 0 dB @ 20 KHz, 60 dB @ 24KHz. Not remotely easy in the analog domain.
Attachments
I'm not a big fan of trying to reproduce analog limitations without good reason. An analyser which notches out or band limits the signal in the analog domain is subject to the limitations of what can be implemented there. If everything is done in the frequency domain and the notching is only for the purpose of calculation I don't see the logic in replicating those limitations, whether that is in the notch or the distortion bandwidth. If your chosen THD+N upper limit is 20 kHz should 22 kHz really only get 30 dB of attenuation? As mentioned above, AES17 only specifies a Q range for the 'standard' notch, there is no specification for shape. Frequency-domain filters are not an issue in general though, appendix B discusses them.
A related question is at what point does the skirt of a fundamental become part of the noise instead of part of the signal? What shape of skirt is acceptable and what shape not acceptable? The AP's notch shape makes that decision for you, with the degree to which the skirt is attenuated depending on its width relative to the AP's notch. What is the theoretical basis for that?
I'm not sure perpetuating the limitations of older equipment is a good thing, but appreciate reproducibility of results across equipment is an issue. However, standards should live up to their name and specify filter shapes if they are considered relevant to a result. Coincidentally the AES17 not-so-standard notch is one of the topics I'll be raising at an AES meeting next week.
In the end it's only calculation, so if folk feel strongly about replicating the AP's notch I can do that, if I can find an accurate definition of the transfer function.
A related question is at what point does the skirt of a fundamental become part of the noise instead of part of the signal? What shape of skirt is acceptable and what shape not acceptable? The AP's notch shape makes that decision for you, with the degree to which the skirt is attenuated depending on its width relative to the AP's notch. What is the theoretical basis for that?
I'm not sure perpetuating the limitations of older equipment is a good thing, but appreciate reproducibility of results across equipment is an issue. However, standards should live up to their name and specify filter shapes if they are considered relevant to a result. Coincidentally the AES17 not-so-standard notch is one of the topics I'll be raising at an AES meeting next week.
In the end it's only calculation, so if folk feel strongly about replicating the AP's notch I can do that, if I can find an accurate definition of the transfer function.
Is Victor´s sine generator really too noisy? A quick check using the Cosmos ADC at 4.5V setting and 48kHz/24bits I get 3uVrms 20-20kHz unweighted noise from the Cosmos ADC (non shorted inputs, stereo mode), and with the latest Victor sine wave generator connected to the Cosmos ADC I get 5.81uVrms, thus Victor´s sine wave generator outputs 5uVrms noise 20-20kHz unweighted.
This highlights several issues with HD measurements. The legacy of measuring with simple analog instruments seems to be driving the design and specs of the standard. With current technology I don't see a lot of value in the THD+N war. Its a difficult goal mostly constrained by the input circuit of the analyzer. And todays state of the art will be able to run into those limits which are mostly noise due to input protection and input scaling requirements. When looking at harmonic distortion separately the floor can be much lower. This also has a valid link to the past, using wave analyzers to measure harmonic amplitudes and adding them together to get a THD number. And with digital analysis more info about the harmonic levels and phase can be had that would be heroic in the analog systems of the past.
While it would be interesting to see if the noise floor increases in the presence of a signal that's not what the analog systems were targeting. A digital system could conceivably separate all stationary artifacts from noise (which could be very useful). The bandwidth of an fft analysis would be defined as much by the window function as any specific mix of bins. Actually measuring noise correctly requires specific knowledge of the bin size and effective bandwidth which comes from the window function used. All this feeds into a useful standard.
While it would be interesting to see if the noise floor increases in the presence of a signal that's not what the analog systems were targeting. A digital system could conceivably separate all stationary artifacts from noise (which could be very useful). The bandwidth of an fft analysis would be defined as much by the window function as any specific mix of bins. Actually measuring noise correctly requires specific knowledge of the bin size and effective bandwidth which comes from the window function used. All this feeds into a useful standard.
I made that osc accordingly posted here schematic and I got THD+N -118db. With another sine source I see -125db on the same ADC. My conclusion is simple - Victor's osc is too noisy for me.
AP SYS2*** implemented that in the DSP, perhaps to match analog vs digital analyzers results.
My understanding, Wien bridge based oscillators all would have high noise level. Filter is too "soft". Another issue, frequency drift.
Tween-T is better choice, https://www.diyaudio.com/community/...range-oscillator.205304/page-487#post-6981973
