How High of a Sample Rate is Enough?

[Blumlein 88]

I need pictures! Can’t someone please produce measurements of a single tone waveform analogue output comparing various inputs from redbook 16/44 then that same signal upsampled to various PCM and DSD rates? Personally I’d like this done in the DX7s as this I the DAC I have

Yes, can do.

But it is a waste of our time. 16 bit will impose some noise limitations with some of the gear. Otherwise, you'll get measurements that are identical effectively.

I happened to have this. Not exactly what you want. 1 khz tone. Done at 48 khz and upsampled the same file to 96 khz. Both were recorded over the same gear at 192 khz sample rate.

Lite blue(cyan) is 96 khz upsampled from 48 khz and magenta is 48 khz at native sample rate. Everything is the same except for one thing. The 48 khz file shows imaging artifacts at 47 khz and 49 khz. The highest being - 116 dbFS. Otherwise more identical than two peas in a pod. It might not look like it, but look close the 1 khz tone and harmonics are right on top of each other. The cyan graph is covering the magenta graph within a fraction of a decibel.

That is all I've seen doing such things. If you are using some odd "audiophile" filter with a reduced treble and slow roll-off allowing more imaging to occur at a higher level, then upsampling will get rid of imaging and restore treble response droops. Of course that is one of those situations where audiophile swears by filter allowing imaging artifacts and softened treble and swears twice as much about the benefits of upsampling which would get rid of those artifacts and flatten frequency response in the treble.

 

[astr0b0y]

Thanks, that’s good to see. I’m also interested in the sine wave reproduction (excuse my non-technical and likely mistaken use of terms) similar to one of the tests Amir does in his DAC test suite. Some DACs have quite lovely and neatly formed sine waves which I interpret as ‘good’ reproduction and conversion from digiital to analog and others show ragged ugly plots that I interpret as a poor reproduction and conversion.
If these graphs I am describing and yours show the same thing in a different format then that’s just one more thing I have learned from ASR.

 

[solderdude]

The ragged sinewaves are scope shots below -90dB the nice looking ones just below 0dB or way above -20dB for instance

You cannot judge the quality of a sinewave by eye. There are small signals riding on it or > 0.1% amounts of distortion that you cannot see on a plot of a sinewave due to the linear scale.
Or you can see some 'ragged' or noisy lines on scopes that are actually showing the limit of the scope (say 10 bit scopes) or the pixels of the picture that is created and are not there in reality.

The FFT plots (as shown above) are FAAARRR more revealing of the actual shape/quality of a sine wave than any scope shot.
One just has to learn to interpret them... but that is true for any measurement.
A single measurement rarely tells everything about fidelity. For that we have a measurement suite and possibly some additional (non standard) measurements.
Proper SQ cannot be concluded from looking at scope pictures unless clipping is seen.
A DAC could be clipping at 0dB but have perfect reproduction at -1dB.

Another format usually doesn't change anything (or very little) about the noise in the audible band as they are artifacts that are not in the file but is noise generated by the components in the analog stage of and behind the actual DAC..

It can be highly educational to listen to sinewaves at different levels where the loudest one is almost piercing and then go down in amplitude in steps of 10dB or so to find out what you can still hear.
There are test discs around or websites where you can simply test this at home.
Don't blow up your speakers. At high levels you should only have a short tone ... and don't go above 1kHz... when using speakers
Start with something like 500Hz.

Knowing about your own audibility limits is FAR more educational than looking at numbers generated by test equipment or hear say and claims of 'I-test-by-ear experts' claiming hearing is superior to flawed and basic tests.

 

[astr0b0y]

These are great explanations from you both, thanks. It’s great to get an understanding of what the measurements mean and what they don’t.
It does make me wonder why upsampling is such a huge topic these days, at least from the perspective of software such as Roon, HQPlayer etc. that all advertise and promote that upsampling will have a positive effect.

 

[SIY]

It does make me wonder why upsampling is such a huge topic these days, at least from the perspective of software such as Roon, HQPlayer etc. that all advertise and promote that upsampling will have a positive effect.

Differentiator in a parity market.

 

[bennetng]

It does make me wonder why upsampling is such a huge topic these days, at least from the perspective of software such as Roon, HQPlayer etc. that all advertise and promote that upsampling will have a positive effect.

Actually it is nothing new. People were using Winamp + SSRC plugin 20 years ago.

 

[bennetng]

I tried @Blumlein 88 's favourite -1dBFS twin tone test on my Realtek ALC892, recorded with a $170 Sound Blaster X-Fi released in 2010. The Realtek is 5dB quieter than my X-Fi but I normalized the waveform before taking the screenshot.

The resampler is from foobar (foo_dsp_resampler mod), it is based on SoX and configured as 95% passband, normal quality and linear phase. I can set it to highest quality but it is meaningless as the differences are not measurable in analog domain. ALC892 is capable of 192k but my X-Fi can only record up to 96k therefore I can only show the spectrum up to 48k.

In case anyone suspect whether my X-Fi is capable of measuring the Realtek or not, I attached my RMAA results and they can be compared with Archimago's ALC892 review:
http://archimago.blogspot.com/2018/05/measurements-msi-x370-sli-plus-am4.html

Based on what a "worthless" Realtek chip and SoX can do, I guess people can determine if their 100-1000x more expensive DACs need another non-free upsampler or not, but I suppose someone will use higher measurement bandwidth and say there are differences above 48k and so on.

 

[Fitzcaraldo215]

Mola-Mola claims near 24 bit and so does MSB

I reckon 22 bit is closer to the actual limit.
The extra bits below 24 bit do nothing as the noise floor remains the same and the signal just gets closer to the noise floor the more attenuation is used.

The volume control (in a practical sense) is valid for 16 bit words in a 24 bit DAC. The 8 bits will actually allow the 16 bit signal to be not compromised when attenuated about 50dB.
As the noisefloor is 'fixed' and not really reducable any 'steps' smaller than 24 bit are simply pointless as they are smaller than the always present noise at the output of the DAC.
So while this works as advertised for redbook in a 24bit DAC it does not apply to 24bit in a 32 bit DAC

I agree that a 24-bit DAC can provide the same effective margin of padding for 16-bit RBCD playback that a 32-bit DAC provides for 24-bit hirez. A digital volume control will be effective in 16-bit padded to 24 without loss of useable signal resolution.

But, I am a little confused by your last sentence. A 32-bit DAC will be just as effective with 16-bit data as a 24-bit DAC on bit shifting volume reductions, though not any more so usably or audibly because of the extra bits. Both will not affect the original 16-bits of signal, which is what we want for digital volume control. Reductions in the "noise floor" seem an insignificant, hypothetical consideration, since the inevitable noise in the original input signal bits themselves will predominate as they are more significant bits than the 8 or 16 extra padding bits. We just don't want the extra padding bits or digital volume control to add any audible noise or distortion, as analog volume controls are inevitably prone to do.

If people are "advertising" a meaningful reduction in noise floor simply by adding the extra bits, it seems useless and inaudible. I think we agree on that. So, back to my earlier premise. Added bits in the DAC, 24-bit for 16-bit playback, or 32-bit for 16- or 24-bit playback, are useful mainly for bit shifting digital volume controls.

Of course, there is another consideration. 16-bit DACs are obsolescent if not fully obsolete, except maybe for extreme low end applications. I expect that newer 24/32-bit DAC chips are more refined and more highly developed in many other ways than just the added bit padding. So, the newer ones might well be capable of also retaining better linearity and a lower residual noise floor to more useable bits in the signal word with competent design. But, they won't deliver 24 or 32 bits free of residual noise, as we agree. We just don't need that many bits for audible quality.

 

[solderdude]

But, I am a little confused by your last sentence

I suppose you mean "So while this works as advertised for redbook in a 24bit DAC it does not apply to 24bit in a 32 bit DAC" this sentence.

What I was trying to explain that while RBCD can profit from 24 bit (and obviously also 32 bit of course) when using digital volume control and all 16 bits of resolution remain 'present' as they don't 'sink' in the noise floor yet (assuming reasonable attenuation is applied) the same won't apply to a 24 bit file when a 32 bit converter is used. That's what I read in your remark about 32 bit and volume control. Reading your reply again I realize I may have drawn the wrong conclusion.
I thought you meant 32 bit is useful because of the extra 8 bits when using 24 bit source files. Of course a 24 bit source file, when digitally attenuated will drown in the noise so won't have any advantages there.

 

[Krunok]

I thought you meant 32 bit is useful because of the extra 8 bits when using 24 bit source files. Of course a 24 bit source file, when digitally attenuated will drown in the noise so won't have any advantages there.

24 bit source file anyhow has lower noise floor than 16 bit file so I don't expect any spectacular "drowning in the noise" effect when volume is lowered below 8 bits capability of 32 bit DAC when processing 24 bit file.

 

[Blumlein 88]

I tried @Blumlein 88 's favourite -1dBFS twin tone test on my Realtek ALC892, recorded with a $170 Sound Blaster X-Fi released in 2010. The Realtek is 5dB quieter than my X-Fi but I normalized the waveform before taking the screenshot.

The resampler is from foobar (foo_dsp_resampler mod), it is based on SoX and configured as 95% passband, normal quality and linear phase. I can set it to highest quality but it is meaningless as the differences are not measurable in analog domain. ALC892 is capable of 192k but my X-Fi can only record up to 96k therefore I can only show the spectrum up to 48k.

In case anyone suspect whether my X-Fi is capable of measuring the Realtek or not, I attached my RMAA results and they can be compared with Archimago's ALC892 review:
http://archimago.blogspot.com/2018/05/measurements-msi-x370-sli-plus-am4.html

Based on what a "worthless" Realtek chip and SoX can do, I guess people can determine if their 100-1000x more expensive DACs need another non-free upsampler or not, but I suppose someone will use higher measurement bandwidth and say there are differences above 48k and so on.

View attachment 14745

So a little imaging at 48 and none at higher rates. Harmonic distortion looks largely the same. 44 has lots of imaging, I bet that Realteck is resampling the 44 with an actual clock for only 48 khz and multiples.

 

[bennetng]

I bet that Realteck is resampling the 44 with an actual clock for only 48 khz and multiples.

Yes, it is written in the datasheet, but the amount of distortion is trivial when compared with early AC97 onboard codecs (e.g. Sigmatel).

 

[Fitzcaraldo215]

I suppose you mean "So while this works as advertised for redbook in a 24bit DAC it does not apply to 24bit in a 32 bit DAC" this sentence.

What I was trying to explain that while RBCD can profit from 24 bit (and obviously also 32 bit of course) when using digital volume control and all 16 bits of resolution remain 'present' as they don't 'sink' in the noise floor yet (assuming reasonable attenuation is applied) the same won't apply to a 24 bit file when a 32 bit converter is used. That's what I read in your remark about 32 bit and volume control. Reading your reply again I realize I may have drawn the wrong conclusion.
I thought you meant 32 bit is useful because of the extra 8 bits when using 24 bit source files. Of course a 24 bit source file, when digitally attenuated will drown in the noise so won't have any advantages there.

Gotcha, and we generally agree. Yes, bits 25-32 are guaranteed to be inaudible below the residual noise floor of any DAC.

But, I still am not totally convinced that a quality 16-bit DAC with a below 16-bit residual noise floor (if there is such a thing) is going to audibly sound much different than a 24-bit with identical noise floor during 16-bit playback, all else being equal, which never happens, of course. Those LSBs below the 16th bit have a hard time being audible in playback almost no matter what.

Ok, you might ask, why then is hirez 24-bit? I think the answer probably has much more to do with headroom on the recording, mixing and mastering side than it does with playback, such the the final levels after up/down adjustment during production remain as noise free as possible. That might not be doable in a 16-bit recording production chain. But, that is pretty much Stone Age these days compared to 24-bit, even in CD production.

So, I still come back to the digital volume control enabled by the added 8- or 16- padding bits as being the key advantage in playback, not necessarily any audible noise floor reduction. Possibly, though, with over sensitive amps and super efficient speakers requiring large amounts of attenuation in our digital volume control, there might be an advantage as you say. But, either a 24-bit or a 32-bit DAC should deliver that quite well with 16-bit playback, and do it better than just about any analog volume control.

 

[dalbert02]

I am the newest guy here, but I'd say 44.1KHz is plenty if implemented correctly. I state this based on Nyquist-Shannon Theorem that states the sample rate need only be twice that of the bandwidth. This has been proven mathematically. I believe even Amir states that somewhere on this site. For further reference, I recommend: http://www.lavryengineering.com/pdfs/lavry-white-paper-the_optimal_sample_rate_for_quality_audio.pdf

 

[Krunok]

I am the newest guy here, but I'd say 44.1KHz is plenty if implemented correctly. I state this based on Nyquist-Shannon Theorem that states the sample rate need only be twice that of the bandwidth. This has been proven mathematically. I believe even Amir states that somewhere on this site. For further reference, I recommend: http://www.lavryengineering.com/pdfs/lavry-white-paper-the_optimal_sample_rate_for_quality_audio.pdf

Heh, it's not that simple actually. Here is the assumption of that theory:
"If a function x(t) contains no frequencies higher than B hertz, it is completely determined by giving its ordinates at a series of points spaced 1/(2B) seconds apart. "

There is really no guarantee that the signal from real life music will satisfy that assumption.

 

[andreasmaaan]

Heh, it's not that simple actually. Here is the assumption of that theory:
"If a function x(t) contains no frequencies higher than B hertz, it is completely determined by giving its ordinates at a series of points spaced 1/(2B) seconds apart. "

There is really no guarantee that the signal from real life music will satisfy that assumption.

But this is guaranteed if a suitable anti-aliasing filter is used between the analogue input and the A/D conversion, no?

 

[Krunok]

But this is guaranteed if a suitable anti-aliasing filter is used between the analogue input and the A/D conversion, no?

And what if there is no analog input and the input is digital as is the case when you're sending the file for playout via USB directly to the DAC?

Generally speaking you have no idea how mastering of that file has been done, right?

 

[svart-hvitt]

Does any audio DAC actually claim, let alone achieve, 24-bit (~146 dB) SNR? Do not know, curious... Note the noise floor seen in an FFT is not the SNR.

https://www.mola-mola.nl/dac.php

Bruno Putzeys aka Kii Three.

 

[andreasmaaan]

And what if there is no analog input and the input is digital as is the case when you're sending the file for playout via USB directly to the DAC?

Generally speaking you have no idea how mastering of that file has been done, right?

Ok sure, we agree I think. This is why in post #3, I suggested 88.2 or 96KHz is safer: these allow for a wider margin of error in applying the low-pass filter during recording and mastering.

Even so, if the mastering engineer has used an incorrect (or no) filter at whatever sample rate they've mastered it at, this will be a problem, i.e. no sample rate is high enough if the engineer uses a sufficiently poor (or no) filter.

And we don't have a problem in the first place if we're using properly mastered RBCD. Filtering at higher sample rates just tends to be harder to stuff up.

 

[dalbert02]

Heh, it's not that simple actually. Here is the assumption of that theory:
"If a function x(t) contains no frequencies higher than B hertz, it is completely determined by giving its ordinates at a series of points spaced 1/(2B) seconds apart. "

There is really no guarantee that the signal from real life music will satisfy that assumption.

Sure there is. How do you think Fourier Transforms work? I humbly suggest reading Dr. Charan Langton's work on Fourier Transforms: http://complextoreal.com/tutorials/#.W3NkpFVKhaT (among her other excellent digital communication tutorials).

Perhaps Lavry's other publication will prove useful as well: http://lavryengineering.com/pdfs/lavry-sampling-theory.pdf