NOS dacs and upsampling

[KMann]

Yes, my big question is, why does non-linearity matter at all at those incredibly low signal levels? I probably have incomplete understanding but my intuition is that this would only affect really quiet passages of music? What am I missing?

Nonlinearity between -140dB and -120dB does not mean the signal has to be very quiet to be affected, what it means is the last 4 least significant bits will be affected. The signal level can be a lot louder, say close to 0dB but small changes at that level that involve the last few bits will be affected.

For simplicity assume 8 bit resolution and for simplicity I am going to consider positive values only. Let us say you get 8 x 6dB = -48dB is the smallest level you can represent and let us say the DAC is not linear below -42dB (i.e. the last bit is not linear)

The smallest value that 8 bits can represent is what is below
[0 0 0 0 0 0 0 1] -> 1/2^-8 -> -48dB, since this bit is bad instead of -48dB it may distort the value to a different level, perhaps -38dB

But, since the last bit is bad, any time this bit is activated, the signal is distorted so even at full scale
[1 1 1 1 1 1 1 1] -> 1 -> 0dB, the last bit is active and will be distorted.

In fact, the signal will distort ~50% of the time!

 

[Rayon]

If we accept the hypothesis that a bit of nonlinearity isn't a big deal up to a certain threshold, but the ear preferring higher bit depth then I find something peculiar: you like 24 bits and 21 but not 22 or 23.

Now from the May measurement graph we can see the linearity is perfect up to 20 bits, but the blue graph with the naked eye does not start to deviate significantly until below -134db, which would suggest 23 bits would be the best compromise.

If you look at the graph closely, the -126db point (21bits) is still very close to 0db. It starts to deviate more right after that. My guess is that for my ears some -0.1db (or whatever it is at 21bit) deviation is still worth it, but -0.5db is already too distorted. My guess on 24bit is that when all those distorted bits are in use, they start to mask each other. I also may have used to the distorted sound of 24. Said that, I currently prefer 21 over 24. My preference hasn't been stable over time, but what I hear, has (more or less).

 

[jamiekp]

ok that makes a lot of sense - thank you.

so back to the top of this thread - ive read different values as being optimal for holo cyan 2 - is 18 the consensus based on measurement?

 

[VNandor]

Goldensound did a nice video on it (the resulting ASR thread was hilarious since they concluded he must have cheated, until Amir came in and declares filter differences are known to be audible)
There's also this nice article: https://www.audiosciencereview.com/forum/index.php?threads/high-resolution-audio-does-it-matter.11/

I've seen the video but not the ASR thread. Goldensound mentioned that he hears above 20kHz which is very surprising for someone his age. It's not surprising if someone hears a sharp cutoff placed in their hearing range. He is just one person it's not really "people" that can reliably pass an ABX test between DAC filters. I assume there are at least some reviewers with enough ego that would love to jump on the opportunity to prove their golden ears. The Bub Stuart paper is interesting although it's done by Bob Stuart of course.

 

[Rayon]

I've seen the video but not the ASR thread. Goldensound mentioned that he hears above 20kHz which is very surprising for someone his age. It's not surprising if someone hears a sharp cutoff placed in their hearing range. He is just one person it's not really "people" that can reliably pass an ABX test between DAC filters. I assume there are at least some reviewers with enough ego that would love to jump on the opportunity to prove their golden ears. The Bub Stuart paper is interesting although it's done by Bob Stuart of course.

Also the interview of PGGB developer in youtube was interesting. He had some theories on why some people perceive non-audible sound: even if we can't hear it directly, the energy needs to dissipate somehow. The high frequency energy can still effect for example the timing of transducers slightly. That could explain it partly. Then I heard about some hairs in human ear that react to high frequency content. We don't hear that as such, but brain still gets information from them and this information is part of brain's interpretation somehow. Some people may have trained their brains so much with critical listening, that this information trigger some Pavlov's dog effect in them

 

[plumpudding2]

I've seen the video but not the ASR thread. Goldensound mentioned that he hears above 20kHz which is very surprising for someone his age. It's not surprising if someone hears a sharp cutoff placed in their hearing range.

While it is indeed unique that his hearing extends above 20khz, in normal music the spectrum in the 20khz range is usually down to -90db, so from a frequency perspective he was able to discern the absence of a -90db signal in an area the ear is very insensitive, masked by 0db music. This is extremely implausible and I hope you agree with that, my one critique of that video is that at the end they offer this as a semi plausible explanation.

The Bub Stuart paper is interesting although it's done by Bob Stuart of course.

Don't knock Bob Stuart just because MQA was a scam, he did some great, peer-reviewed research and got multiple papers published by the Audio Engineering Society that challenged conventional beliefs about digital audio. It's just a shame he tried to make money from it in such a scummy way.

EDIT: Just did a comparison between 8-bit and 24-bit PCM, 1.536 Mhz, LNS15 noise shaper and the long switching time is hampering my ability to do a proper comparison.
(Please someone convince Jussi to implement an instant bit switcher similar to the filter and sample rate switching in the HQPlayer client)
If I made a subjective comment it would be that 8 bits is a bit rougher, 24 bits a bit smoother but it is most certainly not a night and day difference.
Switching to gaussian dither introduces a loud hissing from the noise floor as expected.

 

[Rayon]

And I did some comparison between 8fs with gaussian dither and 8fs noise shaped dither, both from PGGB to 21bit. That gave me same results as what I've got with 20bit earlier. Noise shaped music is smoother and cleaner and imaging is more accurate (I can visualize sounds as clear shapes quite easily). Gaussian dither however feels more intimate and more realistic. With gaussian dither imaging loses micro details and clear borders of distinct sounds are not as clear cut, but those macro elements feel more concrete/real. With gaussian dither music sounds more dynamic, but it can well be that it's the added noise that make those macro elements pop up, which brain associates with dynamics.

It's very clear that noise shaped music carries more information, but it just doesn't make my toe tap. My wild guess is that gaussian dither somehow mimics some noise patterns of real life close enough so that it "decorates" the sound with noise that is familiar and somehow excites ear the right way.

 

[KMann]

And I did some comparison between 8fs with gaussian dither and 8fs noise shaped dither, both from PGGB to 21bit. That gave me same results as what I've got with 20bit earlier. Noise shaped music is smoother and cleaner and imaging is more accurate (I can visualize sounds as clear shapes quite easily). Gaussian dither however feels more intimate and more realistic. With gaussian dither imaging loses micro details and clear borders of distinct sounds are not as clear cut, but those macro elements feel more concrete/real. With gaussian dither music sounds more dynamic, but it can well be that it's the added noise that make those macro elements pop up, which brain associates with dynamics.

It's very clear that noise shaped music carries more information, but it just doesn't make my toe tap. My wild guess is that gaussian dither somehow mimics some noise patterns of real life close enough so that it "decorates" the sound with noise that is familiar and somehow excites ear the right way.

Do you prefer PCM over DSD in your May DAC?

 

[Rayon]

Do you prefer PCM over DSD in your May DAC?

Yes, now that I've been listening to 4fs @ 21/24 bit with gaussian dither. I never liked PCM when I was listening to it @ 32fs/16fs 20bit with LNS15. It sounded somehow lean and boring. For the longest time I was a fan of DSD, but at some point I wanted to start exploring again, trying all kind of things without any prejudices. At that point I was using PGGB and occasionally HQP. I noticed quickly my preferense on gaussian dither vs noise shaped. I also further confirmed my preference of 24 vs the recommended 20. I then started exploring different sample rates. It felt like lower sample rates sounded "taller". They also sound more intimate and "raw". I found out that 4fs was some kind of sweet spot for me with 24bit. Sound is already much more open vs 1fs, but I still have this intimate connection to music. 2fs sounds still kind of "dry", but 4fs already creates this disconnection with me and the music. It's like the higher sample rate can paint the picture, but suddenly that picture is also in front of you, instead of touching you. 4fs keeps me more alert.

I then experienced a bit further with bit depth, keeping sample rate fixed and found out that 21 bits already solved the lean/boring sound of 20, but it doesn't introduce too much noise yet.

So in case you want to try things out and see if this is relatable to you, relevant things to remember:

1. Use 21 or 24 bits
   • If I listen to 20bits, I keep skipping songs constantly and seeking different spots, until I realize that it's in the system, not track's fault
2. Use 4fs, but try also 8fs
   • Higher sample rates paint the picture instead of you being within the picture, part of it
   • 8fs already creates that distance, but I occasionally use it as a middle ground
   • When listening to 4fs, listen to transients
     • I know that higher sample rate should give us better timing, but to me the transient response of 4fs is better/snappier while higher rates lose that snappiness
3. Use gaussian dither
   • With 8fs also noise shaping can be tried out, but with 4fs it's just horrible IMO (at least in PGGB)
   • When switching between gaussian dither and noise shaping, remember to give them both a change
     • When doing quick a/b, the benefits of each don't come up
     • One needs to live with both for days and then do the comparison
       • When doing the comparison, listen to what happens in the upper registry
       • Gaussian dither sounds more dirty when doing quick a/b, but after a while, you don't notice it anymore, but you're left with sound that excites your ears and keeps your toe tapping
         • Gaussian dither also does this in very natural way
         • I don't get the same effect with TPDF

My recommendation is to live with 4fs/21bit/gaussian for a week exclusively and see where it takes you

 

[plumpudding2]

When switching between gaussian dither and noise shaping, remember to give them both a change

• When doing quick a/b, the benefits of each don't come up
• One needs to live with both for days and then do the comparison
  • When doing the comparison, listen to what happens in the upper registry
  • Gaussian dither sounds more dirty when doing quick a/b, but after a while, you don't notice it anymore, but you're left with sound that excites your ears and keeps your toe tapping
    • Gaussian dither also does this in very natural way
    • I don't get the same effect with TPDF

Completely agree with the toe-tapping effect of Gaussian dither. Also I agree that for subtle system changes it is best to live with it for a few days and then switch, your brain needs time to adjust to a certain system and when the switching happens you have a small reacclimatization period to notice changes again.

Yesterday I just wanted to do quick switching since I thought the effects of 8-bit would be horrible. I managed to do it better, quick-switching between PCM and DSD as a reference and then changing 8 between 21 bits. Then it very quickly became obvious that 21 bit PCM sounds very similar to DSD, a bit flatter and less detailed but with more immediacy, while 8-bit PCM even with noise shaping loses 70% of its dynamics.

 

[Rayon]

Then it very quickly became obvious that 21 bit PCM sounds very similar to DSD, a bit flatter and less detailed but with more immediacy, while 8-bit PCM even with noise shaping loses 70% of its dynamics

This very nicely summarizes what I hear and also touches the trade off I see between this and DSD. It's like DSD "takes you there", while this "brings it to where you are", especially with 4fs. I got inspiration to this when listening to some 96/24 albums as-is, such as Chick Corea - Plays (2020) [24B-96kHz]. It was like the recording situation happened in the room where I'm in. It's very life-like experience, while it's not completely clean. Upsample it to DSD and you get picture of the space and situation, but your your body is no longer connected to the music same way. I have now l listened to >20bit 4fs exclusively for at least a month. I do occasional visits to higher sample rate PCM and DSD, but I always come back to this. It seems to be my pareto optimum.

TA-ZH1ES teached me something very important. It sounded great. It was clean, it had big soundstage (at least at my standards at the time) and so on. But I didn't listen to music anymore. I then was reading some blog post or thread post and the person commenting mentioned something about the product: "the toe must tap". I realized that this was the problem. I was just listening to the scenery. I occasionally cried on some tracks (that's important for some). But most of the time, I was bored to death. I tried some things and ended up buying Qutest. Qutest was technically inferior in terms of sound stage and clarity, but it did something that TA-ZH1ES didn't do for me: the sound was exciting and my foot was moving again.

4fs 21bit with gaussian dither is doing the same thing philosophically. It's not the biggest stage, it's not the cleanest sound and it doesn't make me smile as often as DSD does. But it drives me. I constantly notice that I'm enjoying music and music guides me.

 

[KMann]

This very nicely summarizes what I hear and also touches the trade off I see between this and DSD. It's like DSD "takes you there", while this "brings it to where you are", especially with 4fs. I got inspiration to this when listening to some 96/24 albums as-is, such as Chick Corea - Plays (2020) [24B-96kHz]. It was like the recording situation happened in the room where I'm in. It's very life-like experience, while it's not completely clean. Upsample it to DSD and you get picture of the space and situation, but your your body is no longer connected to the music same way. I have now l listened to >20bit 4fs exclusively for at least a month. I do occasional visits to higher sample rate PCM and DSD, but I always come back to this. It seems to be my pareto optimum.

TA-ZH1ES teached me something very important. It sounded great. It was clean, it had big soundstage (at least at my standards at the time) and so on. But I didn't listen to music anymore. I then was reading some blog post or thread post and the person commenting mentioned something about the product: "the toe must tap". I realized that this was the problem. I was just listening to the scenery. I occasionally cried on some tracks (that's important for some). But most of the time, I was bored to death. I tried some things and ended up buying Qutest. Qutest was technically inferior in terms of sound stage and clarity, but it did something that TA-ZH1ES didn't do for me: the sound was exciting and my foot was moving again.

4fs 21bit with gaussian dither is doing the same thing philosophically. It's not the biggest stage, it's not the cleanest sound and it doesn't make me smile as often as DSD does. But it drives me. I constantly notice that I'm enjoying music and music guides me.

Interesting, to me it looks as though you prefer lower rates due to better dynamics (hence the toe-tapping), though I do not know why lower rates should sound more dynamic. Do you use any compression (wavpack or FLAC)?

 

[Rayon]

Interesting, to me it looks as though you prefer lower rates due to better dynamics (hence the toe-tapping), though I do not know why lower rates should sound more dynamic.

One theory I've come up with is that with higher rates you can better hear the rise of the transient, while with lower rate it's already on the way up before your ear catches it. Hence it may appear as dynamic/aggressive. But no idea either.

Do you use any compression (wavpack or FLAC)?

Currently I've been using FLACs, but havent heard any difference to WAV.

 

[gregorio]

There has been quite a bit discussion on upsampling in the context of NOS dacs, especially R2R.

Yep, bizarre isn’t it? They buy a Non-OverSampling DAC and then oversample it.

Often for oversampling dacs the highest sample rate and bitrate that the dac can accept sounds best

Nope, that’s pretty much never the case. In fact technically, the highest sample rates are worse.

Examples of questions that have been on the table several times:

• [1] What bit depth should one choose when using PCM
  • [2] Should one go with the lowest bit depth with perfect linearity or is there something to be gained if some non-linearity is tolerated to get higher bit depth
• [3] What sample rate sounds best
• [4] PCM vs DSD and why

1. When choosing bit depth it doesn’t matter. Each bit represents about 6dB of dynamic range (6.02dB to be precise), so 16bit has about 96dB dynamic range. Music recordings typically have less than 50dB dynamic range, only about 30dB dynamic range with the highly compressed recordings, up to about 60dB with some uncompressed orchestral recordings and a very few classical recordings that go to about 70dB dynamic range. Studio/Music microphones only typically have 70dB or less dynamic range and the most dynamic go up to just over 80dB. In all cases, this is a lot (or a massive amount) less than the dynamic range offered by 16bit, even the highly dynamic (60dB dynamic range) orchestral recordings are effectively using just 10bits of the 16bits, the remaining 6bits are just random values (noise). So, what does 20bit or 24bit get you except ever quieter levels of noise? 24bit is useful when recording because it allows a huge amount of headroom but for playback there is literally nothing (other than noise) to be gained.

2. Not quite sure where this one came from. The lowest bit depth has less linearity but once you get beyond a few bits then dither eliminates any non-linearity, hence why it’s a requirement of digital audio.

3. 44.1kHz or 48kHz. If you go above that, then you run the risk of ultrasonic content, content that is inaudible to humans but may cause an audible non-linear response (inter-modulation distortion) downstream in your amp or transducers. This is not an issue as far as oversampling 44.1 or 48kHz on the DAC end is concerned, as an anti-image filter is applied to remove everything above the Nyquist frequency (half the sample rate).

4. Technically PCM, as DSD inherently has a distortion issue due to only having 1bit and therefore cannot be dithered appropriately. However, in practice (with music) this is pretty much never an audible issue, so it doesn’t really matter which you choose. 16/44 PCM uses less bandwidth than DSD but historically, SACD (DSD) versions were sometimes mastered better than the 16/44 version, as it couldn’t be ripped and could only be played in good listening environments. So, there’s no golden rule, it depends on the mastering of a particular track/album.

That's the problem, according to science nothing matters below ~-110db and above ~15khz depending on age, yet people can reliably pass ABX tests between dac filters that only affect 20khz and things like noise floor modulation are very audible despite being below these limits.

That is not “according to science” and yes, noise floor modulation is very audible below these limits, if in fact they’re not below those limits, for example if you choose a quiet section of music and raise the playback level significantly (as science indicates). At a reasonable/loudish playback level, say 85dBSPL peak with HPs for instance, then -110dB would be -25dBSPL, which cannot even exist as a sound pressure level and therefore obviously has to be inaudible.

I suggest moving this discussion over to the Sound Science forum if the OP or anyone else wants the actual facts, as discussing science isn’t allowed in this or any other forum on Head-Fi except the sound science forum. If the answer you’re looking for is just marketing misinformation or audiophile myths based on marketing misinformation, then it’s fine to just leave it here.

G

 

[plumpudding2]

4. Technically PCM, as DSD inherently has a distortion issue due to only having 1bit and therefore cannot be dithered appropriately. However, in practice (with music) this is pretty much never an audible issue, so it doesn’t really matter which you choose. 16/44 PCM uses less bandwidth than DSD but historically, SACD (DSD) versions were sometimes mastered better than the 16/44 version, as it couldn’t be ripped and could only be played in good listening environments. So, there’s no golden rule, it depends on the mastering of a particular track/album.

Is this statement based on that old Vanderkooy & Lipschitz paper? There's a reason most dacs convert to DSD internally, with any multi bit solution you either get settling time issues or ultrasonic aliasing around the Mhz range. All their paper proved is you can't naively dither DSD or overloading will occur.

That is not “according to science” and yes, noise floor modulation is very audible below these limits, if in fact they’re not below those limits, for example if you choose a quiet section of music and raise the playback level significantly (as science indicates). At a reasonable/loudish playback level, say 85dBSPL peak with HPs for instance, then -110dB would be -25dBSPL, which cannot even exist as a sound pressure level and therefore obviously has to be inaudible.

Yes Benchmark uses 0dbspl as the threshold of absolute inaudibility and most music when played back does not have peaks above 110db. Of course you can contrive a situation where these effects become audible, but my point was that generally speaking these effects should not be audible.

I suggest moving this discussion over to the Sound Science forum if the OP or anyone else wants the actual facts, as discussing science isn’t allowed in this or any other forum on Head-Fi except the sound science forum. If the answer you’re looking for is just marketing misinformation or audiophile myths based on marketing misinformation, then it’s fine to just leave it here.

G

Discussing science isn't allowed? The only thing isn't allowed is telling cable buyers etc. that everything they hear different is imagined, to protect threads against pointless rehashed debates and mud slinging.
A thread where people constructively discuss how their objective and subjective experiences diverge and might possibly be reconciled is the exact opposite, unless sound science forum users start leaking out and proceed to tell everyone they're wrong..