You have tested with all dac architectures? IIRC, @MarcelvdG said that for some architectures as little as a few ps could be significant. Not enough to produce bit errors (like people look for in CD players), but possibly enough to produce other types of artifacts.
Also IIRC, @EdGr did a calculation of what time errors would be significant at 24-bits (which some dacs are now getting closer to). Again IIRC, it was down in the low ps region.
Maybe, possibly, could be.
Of course I have not made a life testing every possible combination. No one has. Let's look at the reality of it, and you are right now picking on one specific aspect that supports your position "that we can't know everything". But, we know enough.
As I asked before, what peak time difference is there in your jitter? Compare that with the sample rate, sample depth has no bearing on your question. We are talking about time here. Now, compare that with the HF cut off of the filters used in DACs. Let's support your argument for a moment, say we have a one interval error. What is the frequency this represents? Okay, both channels used this common clock so both channels are offset in time by the same amount. Humans hear time or phase differences between time of arrival between our ears. If that time differential occurred both ears are offset by the same amount. Hmmm, we could never detect it.
It would take gross errors in decoded amplitude over longer time spans for us to actually hear it as distortion. The high sample rates mean errors like these are extremely high in frequency, and due to filtering would be extremely low amplitude. Then there is our hearing to contend with.
In test results we can see far more than human bodies are capable of detecting these days. I see anomalies that are artifacts of noise, cable and other test setup issues. These are normally way below our threshold of sensing in any way. But, there they are on the screen. Their existence doesn't support your view. In fact, I am amazed at what people don't hear that I can readily see and reproduce. Basically, if you can sense something in any way, it is there like a sore thumb with some of today's test equipment. Not everyone can interpret these things, but they are there to be seen and measured.
All I have ever said to you Mark, is that you need to become familiar with these tests and use of equipment and apply this to what you are saying. You will find the truth following this path. It isn't faith, it is experience and reproducible tests. The audio system on my bench is pretty good. Klipsch THX 6000 speakers (99dB/watt), decent electronics: Marantz 300DC modified (to reduce distortion), Yamaha C50 preamp and various sources. I need to hear everything along with measuring it. I do listen carefully. My other systems use PSB Stratus Gold speakers and Klipsch RP8000 speakers. Maybe not the best, but you get a good feel for how things perform. I listen to everything on the market once it hits the bench, and have for many decades. I correlate both measuring and listening.
So I am not a guy who watches a needle. I look at response spectrums, residuals from THD meters and have used very good equipment to do so. Then, like everyone else here, I listen. I love music, it's always playing here. If something is off, it is noticed and chased down.
Of course I have not made a life testing every possible combination. No one has. Let's look at the reality of it, and you are right now picking on one specific aspect that supports your position "that we can't know everything". But, we know enough.
As I asked before, what peak time difference is there in your jitter? Compare that with the sample rate, sample depth has no bearing on your question. We are talking about time here. Now, compare that with the HF cut off of the filters used in DACs. Let's support your argument for a moment, say we have a one interval error. What is the frequency this represents? Okay, both channels used this common clock so both channels are offset in time by the same amount. Humans hear time or phase differences between time of arrival between our ears. If that time differential occurred both ears are offset by the same amount. Hmmm, we could never detect it.
It would take gross errors in decoded amplitude over longer time spans for us to actually hear it as distortion. The high sample rates mean errors like these are extremely high in frequency, and due to filtering would be extremely low amplitude. Then there is our hearing to contend with.
In test results we can see far more than human bodies are capable of detecting these days. I see anomalies that are artifacts of noise, cable and other test setup issues. These are normally way below our threshold of sensing in any way. But, there they are on the screen. Their existence doesn't support your view. In fact, I am amazed at what people don't hear that I can readily see and reproduce. Basically, if you can sense something in any way, it is there like a sore thumb with some of today's test equipment. Not everyone can interpret these things, but they are there to be seen and measured.
All I have ever said to you Mark, is that you need to become familiar with these tests and use of equipment and apply this to what you are saying. You will find the truth following this path. It isn't faith, it is experience and reproducible tests. The audio system on my bench is pretty good. Klipsch THX 6000 speakers (99dB/watt), decent electronics: Marantz 300DC modified (to reduce distortion), Yamaha C50 preamp and various sources. I need to hear everything along with measuring it. I do listen carefully. My other systems use PSB Stratus Gold speakers and Klipsch RP8000 speakers. Maybe not the best, but you get a good feel for how things perform. I listen to everything on the market once it hits the bench, and have for many decades. I correlate both measuring and listening.
So I am not a guy who watches a needle. I look at response spectrums, residuals from THD meters and have used very good equipment to do so. Then, like everyone else here, I listen. I love music, it's always playing here. If something is off, it is noticed and chased down.
Would you enlighten us on how a signal can fall between two adjacent FFT bins? I've always been of the impression that where one bin ended, another began with zero overlap. That would make the probability of a signal falling exactly at the bin boundary equal to zero. After all, the integral of the probability density function from A to A is zero by definition.
Yep. Did that back in college. We did convolution the same way. That was a while ago, though. 🙂
Tom
Looking at the plots above, I'd guess that the raised noise floor on some of them (e.g., the green trace on the one directly above) is caused by HF junk (> than the ADC bandwidth) in the measurement loop. Clearly, the isolator is doing a good job on the ones with the >-140 dB noise floor. Another trick you may want to try to get more depth on the noise floor is clamp-on lossy ferrites on the USB cables.
Hi all,
I haven't followed this thread and I don't think I want to either, so I don't know if what I am about to write is relevant to the discussion, but to give you some context to Mark's reference to something I wrote:
When you have a sigma-delta type of DAC, you by definition have a heavily oversampled and coarsely requantized digital signal going into a coarse DAC. Due to noise shaping, it has almost no audio noise but lots of out-of-band quantization noise. A bit of jitter on the DAC clock, jitter that is equivalent to far-off phase noise, can convert some of that out-of-band quantization noise into the audio band and mess up the audio noise floor.
To what extent this happens depends on the DAC type. Sometimes a few picoseconds is enough for a substantial audio noise increase. When playing silence at a volume setting high enough to hear the noise, it will be audible.
Systematic jitter due to clock spurs around half the sample rate is even worse, if the sigma-delta modulator has strong tones around half the sample rate, like straightforward single-bit modulators normally have. You can get all sorts of funny artefacts at low audio signal levels from that.
Regards,
Marcel
Hi Marcel,
That, I can see.
However, it is a special case and doesn't apply to amplifiers and other sources. I see your point as perfectly valid as an example of a problematic DAC design, one specific type. It would also show up clearly on your test bench, and once that fault was resolved due to the ability to measure it, no longer an issue. So once again, our ability to measure will isolate and resolve an issue that was audible.
My issue is the derailing a discussion using a specific hypothetical idea that applies only to a special case. Something unproved and not really relevant. If it was something we could hear but not measure - okay.
That, I can see.
However, it is a special case and doesn't apply to amplifiers and other sources. I see your point as perfectly valid as an example of a problematic DAC design, one specific type. It would also show up clearly on your test bench, and once that fault was resolved due to the ability to measure it, no longer an issue. So once again, our ability to measure will isolate and resolve an issue that was audible.
My issue is the derailing a discussion using a specific hypothetical idea that applies only to a special case. Something unproved and not really relevant. If it was something we could hear but not measure - okay.
SOTA DACs have surprisingly little out of audio band noise, it is incredibly well suppressed. Visible noise shaping was an issue years ago, but it is not a case anymore.
Last edited:
Nope. Only for a continuous time FT. When we do a DFT it is an approximation, and frequencies in between bins have partial correlation with adjacent bins.
Just try this: Google the following, "an fft (or dft) gives a measure of signal correlation with bin frequencies" and see what you get. You will find that it is a fact.
Also, try reading the first and third paragraphs here: https://www.dspguide.com/ch8/6.htm
Last edited:
Incorrect. For small errors, an error in time is just as significant as an error in amplitude. They are considered more or less equivalent errors. Thus in a 24-bit dac we need to be just as concerned about PN as we are about AN. Resistance to time errors that are far out from the carrier (as viewed in the frequency domain) is what modern dacs are good at. Errors closer-in to the carrier (the nominal clock frequency) tend to show up in the audio band. Yes, they can be turned into noise by scrambling up low level information, but there goes ITD imaging, which nobody is checking by measurements because they don't know how to measure it with their AP machine.
Also, regarding RFI/EMI causing veiled or grainy sound in amplifiers, as claimed by Bill Whitlock, it is caused by demodulation/remodulation effects in semiconductor junctions. This stuff is not new at all. Every EE in audio should understand how that works. Also, none other than @1audio has claimed that EMI/RFI susceptibility is one of the things we need to do more work on. Not just PSS HD/IMD and noise floor, which he (1audio) likened to a drunk looking for his car keys under a streetlight because the light was better there.
Also, regarding all those FFTs recently posted showing the effects of USB isolators (despite the curious worsening of some signal aspects). How many people are using isolators with their USB dacs? My dac uses galvanic isolation at the I2S level to accomplish the same type of benefit, but it does it closer to the dac chip. And, yes the difference of isolating or not is easily audible. The worst case is when the USB board is USB powered. Awful. But it can look very nice in a cleanly powered FFT analyzer such as a 555x.
Last edited:
All of the plots I posted are of different test system configurations. Each configuration is described in the text just above each plot. In addition, the set-up is shown in the comment header at the top of each plot.
Each plot has two traces. The black trace is the spectrum averaged 16 times, as displayed in the lower left corner of each plot. The red trace is the captured peaks for that same measurement. I let the peak capture run for about 30 seconds, although I did not sit there with a stop watch to time it. The legend for the traces are displayed at the bottom of each plot in the footer.
My only goal in posting these was to illustrate, with actual measurements, how systems tend to have unwanted currents running about that can cause unwanted signals in the audio band.
Each plot has two traces. The black trace is the spectrum averaged 16 times, as displayed in the lower left corner of each plot. The red trace is the captured peaks for that same measurement. I let the peak capture run for about 30 seconds, although I did not sit there with a stop watch to time it. The legend for the traces are displayed at the bottom of each plot in the footer.
My only goal in posting these was to illustrate, with actual measurements, how systems tend to have unwanted currents running about that can cause unwanted signals in the audio band.
So I did that. Here's what Gemini says:
"Correlation Measurement:
The magnitude of each complex number in the output represents the strength of the signal's correlation with the corresponding frequency bin. In other words, a large magnitude for a particular bin indicates that the signal has a significant component at that frequency. "
That's congruent with my understanding of FFTs and DFTs. I'll just lump the two together and call them FFTs from now on, okay? An FFT is 'just' an optimized way of computing a DFT. It turns out that you end up redoing a lot of the same math when you compute a DFT. FFT removes the redundancy. The end result is the exact same.
I then Googled "fft signals falling between bins". I will openly admit that I don't understand each and every word. The math gets pretty hairy. But the impression I get – which is also congruent with what I learned in college – is that all the energy in the signal is represented in the FFT bins. Otherwise, the FFT would violate Parseval's Theorem. In other words, no signal falls between the FFT bins.
Of course, you can have a scenario where two frequencies fall really close together. If you use a short FFT length they'd be lumped together into the same bin, so you lose frequency resolution (but the FFT is faster and requires less memory). If resolving frequencies that are very close together is a concern all you have to do is increase the FFT length (or decrease the sampling frequency). For example if you use a "1M" FFT (1.049M-sample = 1024*1024 samples) you get a frequency resolution of 42 mHz (= 0.042 Hz).
Tom
Hi Mark,
Here we go again. If your bit depth is 16 bits, you'll have a very similar analogue amplitude to a 24 bit or 32 bit sample. I do agree, the time error is what is significant, not the analogue amplitude, which will be similar in amplitude to adjacent samples. I did address this specific point earlier. You haven't responded.
We know all about audio rectification and everything else. Now you have derailed the discussion from audio analyzers into a focused attempt to debunk the value of measuring things. Oddly enough, you are referring to these measurements, meaning they show up.
How about we return to the topic of this thread. You can begin a thread on your interests.
Here we go again. If your bit depth is 16 bits, you'll have a very similar analogue amplitude to a 24 bit or 32 bit sample. I do agree, the time error is what is significant, not the analogue amplitude, which will be similar in amplitude to adjacent samples. I did address this specific point earlier. You haven't responded.
We know all about audio rectification and everything else. Now you have derailed the discussion from audio analyzers into a focused attempt to debunk the value of measuring things. Oddly enough, you are referring to these measurements, meaning they show up.
How about we return to the topic of this thread. You can begin a thread on your interests.
No, I was responding to multiple claims in posts made after my last post. It is other people who keep trying to find holes in what I am trying to explain instead of trying to understand it.
Getting back to close-in clock phase noise, it is like AN voltage regulator noise. It is timing error offset from the carrier by a difference frequency in the audio band. Therefore it is changing at audio band rates, and sequences of samples are affected in the audio band. Another way of thinking of it is that the clock frequency is drifting around at audio frequency drift rates.
Moreover, many jitter measurements and jitter specifications don't show it because they assume such slow clock drifting shouldn't be considered to be defined as jitter. Often jitter is calculated from phase noise measurement where phase noise is not counted as jitter until it is 12kHz offset from the carrier. That means it would be counted as jittering the audio only at 12kHz and above.
Last edited: