Marvey's DAC Chart of Awesomeness

Just plotted the spectrums of the two files to see how much HF noise the quantization added. As expected, the difference is minimal and it's very unlikely that this is the cause for the audible differences between the two files.
The two are nearly the same up to 46kHz (or maybe even higher). Beware of the two slightly different scales, Audacity changed the scale because of the slightly higher noise with the 16 bit file. However there are some changes to the actual shape of the curve, even in the audioband but I'm not sure if looking at individual pixels in the two images is the right thing to do.

24 bit:


16 bit:


Additionally I listened to the difference file at a volume level that I realistically listened to ...
(I listened to it a pretty high volume - at about 2 o'clock on the Ragnarok volume put at mid gain (say about 18 steps from no attenuation, I believe each step is 1.25db), Gungnir XLR output (4V), Rag XLR output into HD800 (use innerfidelity specced sensitivity). Feel free to calculate the maximum (0dbfs) volume level if you want. FYI the music file at the position that I listened to it is about -20dbfs (which should make quantization error much more audible), so no, I didn't blow my ears) ...
and I think I can now hear the added noise floor over the ambient noise floor. I actually want someone to do the math for me. I have a feeling that I would fail.

(Actually, let me try: 4Vrms * (Gain = Rag mid gain = 5 = 14db - 18*1.25db)= -8.5db = 4Vrms * 0.375837 (?) = 1.503348 Vrms
90db = 0.242 Vrms at 500Hz (HD800 unmodded)
1.503348Vrms / 0.242Vrms = Factor of 6.711375
Factor of 6.711375V == +16.53623db (?)
That would put the maximum volume at about 106.5db. 106,5db-96db = 10.5db. Seems about right given the ambient volume, etc. I also tried to see if I can hear a -90db 1kHz sine. I can very easily hear it. The sine seems much louder than the difference file.
Someone please check if I made a mistake somewhere. I have a feeling that I failed - my REW volume calibration is much quieter, so either the HD800 has awesome low distortion of less than 0.1% at 106db, or my math is off (or the innerfidelity data is off).)

LOL I just realized something. For some reason I had digital volume calibration in windows enabled -> -14.4db !!! I will still leave the above part in here. I listened to the files in Audacity (to listen to the difference). Once I opened the files in foobar (where I'm using ASIO output and no digital volume control), I jumped because it was so much louder (it wasn't super loud, I just jumped because it was much louder than before). This of course means that the maximum level was 14.4db quieter than my math above suggests (about 92.1db at 500Hz) and the white noise was 96db below that. This also means that my volume calibration in REW is pretty darn accurate.
92.1db - 90db = 2,1db -> The -90db 1kHz sine was probably between 0db and 5db. I could easily hear this.
92.1 - 96db = -3.9db -> This is how loud the difference file was. Of course there are some inaccuracies, so could be slightly louder. I think I can hear this. My room is very quiet.​

I tried it now. The difference is similar to the one that I heard with my dithered file. I can hear the difference in a blind test.

EDIT: Another thought: With the 18 bit file I might actually be getting into where my Gungnir might become the limiting factor. The difference between the 16 and the 18 bit file certainly seems to be much larger than the difference between the 24 and the 18 bit file.

 

The gear I used was an ATH portable dac/amp and K7XX+EQ, and a random hi-res sample from 2l test bench IIRC. I may try it again with better gear when I get back home next month. Also I might be having expectation bias since the difference waveform is inaudible.

Some thoughts: If we assume dBFS scale is proporsional to dB SPL, +-(-96dBFS) noise on -50dBFS hiss would result in roughly +-1/34 perceived loudness at max, and is audible(?). Perhaps the whole inaudible extracted difference = inaudible difference concept is wrong. Would be interesting to see the what happen if -96dB white noise/sine wave is added to the 24bit sample intentionally. It's almost 7am maybe what I said doesn't make any sense idk.

 

That's what I think. But as I said, at night time it seemed like I could just barely hear the hiss of the difference file.

I thought about that too. Will try that now. I wonder if dithering randomizes the quantization error enough, because technically it should do just that (simply add noise to the file, but not white noise). Maybe the dithering in Audacity is simply not good enough, or maybe what I heard was just a result of the slightly higher noise floor.
Personally I have a hard time believing that. Masking should prevent me from hearing noise at such a low level while music is playing, even more so if it seems to be just around the audibility threshold.
I'm going to try that now. White noise with an amplitdue of 0.00002 seems to be just slightly louder than -96db. The noise of the difference file also seemed slightly louder than -96db. I'm going to go with that now. The noise is 24bit so should be "more random" than the quantization noise.

Will update in a few minutes.
Update:
The noise that I added to the file actually turned out the be slightly louder than the quantization noise. I think it's too early for conclusions but I don't think I can tell them apart (the original and the 24bit one with added noise). Adding white noise (even if slightly louder) does indeed seem to be more benign than quantization error. If anything the difference is similar to the 18bit vs 24bit, even if the difference with the 18bit file would be much quieter than the 24bit file. Maybe the file with added -92db white noise actually sounds better than the 18bit file.

My own preliminary conclusions:

• Undithered 16bit files are not audibly transparent. Some people say that it will be dithered sufficiently by noise on the recording. This doesn't seem to be true. The recording had noise multiple orders of magnitude larger than the quantization noise and I could still tell the difference.
• Dithered 16bit files are also likely not audibly transparent. I've not gone through every magical option of dithering but the ones that I've tried sounded different than the original 24bit file.

TBD:
- Is adding white noise at -92db audibly transparent?
- Does adding 24bit white noise to the 16bit file make it sound better? Does adding 16bit -90db white noise sound worse than 24bit white noise?
- At what bit-depth does the quantization error become inaudible?

I think I read somewhere that the Gungnir has an ENOB of around 18. The Gungnir MB has 19bits of resolution. Ideally I'd want to use an Yggdrasil or similar for tests with files of 18bits or more.

 

Can someone try running a 192khz version of a song and a 44khz version through Bifrost Multibit/Gungnir Multibit/Yggdrasil and compare those?
And then try resamplers from players like Audirvana, HQplayer,etc... Just to see which of the resamplers sound the closest to the original? I know resampling can get really subjective (pre/post ringing, linear/minimum phase,...) but the consensus at SBAF seems to be that the Schiit resampler is the best there is and I'd like to know how close it can get to the real 192khz file.

 

Since the quantisation error is not random and white noise is, 15 bit white noise should not have an effect less than 16 bit quantisation noise even in the worst case scenario(?). Maybe try white noise/wave pattern at -90dB/-102dB to compare to 16bit/18bit. I honestly doubt that adding -144dB noise to anything would make a difference, but I might be wrong. I think schiit DAC's ENOB is just the rated SNR/6.02, but if "inaudible extracted difference = inaudible difference" is false, then ENOB doesn't do anything here(?).

The best I found with foobar SOX resampler is best quality, linear, allow aliasing from inspecting waveform difference. Didn't bother to do ABX with different settings tho.

 

Good question. Technically a file with -90db 24bit white noise added should probably not sound better than at least the dithered 16bit file. I will try files with 24bit and 16bit added white noise.
I now really do think that quantization noise really isn't comparable to normal random noise, as tends to be assumed by many. Quantization noise might be much nastier than regular white noise and have a greater overall impact on fidelity. The quantization noise for the files sounds pretty nasty.

Technically, I believe that is how ENOB is usually specced, but I also think that there's more to it. Like Marv said, multibit DAC chips with higher INL/DNL will have a lower ENOB, even if this isn't taken into account for ENOB specs.
But yes, if the DAC chip could technically resolve the information and it is limited by noise elsewhere in the circuit, then the difference could maybe still be audible, or something. I'm really not so sure about this theory. Technically "inaudible extracted difference = inaudible difference" should always be true, which is also why I listened to the difference file again. With the 16bit file the difference file was just over the audibility threshold. Hard to believe that differences are really caused by that, but not impossible.


On another note: I might have been wrong in assuming that there was no music information baked in my difference files. I just took a second look at the files. Some files have it and some files don't. I'm a little confused right now as to which one have it and why and which ones I used for the comparisions. I believe that my undithered 16bit file was fine, but maybe the dithered one wasn't, or maybe it was. I don't really know right now. I still need to try more things to see how I got get rid of that. Your 'asdf' file doesn't have any music information. I'm actually not too concerned about the music information in the difference files as I could also tell the difference between the original 24bit file and your dithered 16bit file. In other words, I don't think that it makes my results invalid, but I will have to look into it. The difference file between your 'asdf' file and the original file also doesn't sound like random noise. Maybe there is indeed a special type of dithering that sounds better than others.


I also feel like this discussion should continue in a seperate thread. I essentially hijacked and killed this thread. Maybe someone can move all the posts related to this subject to a new thread, or maybe not.

 

Whatever waveform I throw at it below -96dBFS is inaudible through my system and environment, even when I crank everything up, so I will just skip ABXing these files.

It is not random indeed, the adsf file is dithered and the noise increases along the frequency, but the noise is consistent in the time domain.

Here's what I got after amplifying all the noises. The 16bit non dither difference file sounds and looks the same to white noise even in a spectrogram, and yet being nastier than white noise, this is what bugs me. The difference noise in asdf is the 24-16d file, consistent with time, and sounded much nicer than white noise IMO.

 

I have a Rag, so plenty of power. I can crank the -96db files up very loud.
At higher listening levels the -96db file should be just audible in a very quiet room. I can hear it at night, but not right now. Even then, could very well be that you don't have enough power.

You're right. I have both of them side by side now. The spectrum overall is identical. They also seem to sound pretty much identical (the difference files).
Will have to check if I can hear a difference with the added white noise again. Maybe all of this turns out to be BS.

EDIT: There is certainly a difference between -90db 16bit noise (1bit) and -90db 24bit noise (way more bits). I think this is what I meant above. Quantization essentially creates the 1bit type of noise. Normalized they're different volume levels because of how the waveforms differ. The 16bit file is less loud. I think they also sound slightly different.

EDIT: The two files with the two different added noise files sound different. I actually feel like the one with the added 16bit noise might be closer to the original, which should make sense if the 16bit 0.000019 amplitude noise is indeed less loud than the 24bit 0.000019 amplitude noise.
All of this seems to be much more complicated and more confusing than I thought it would be.

 

Two LPs I noticed recently that show vinyl's lack of high frequency dynamics and transients:
Angel Witch self titled original 1980 LP vs the 2010 Sanctuary 2x CD
Morbid Angel - Altars of Madness original 1989 LP vs the 2016 Full Dynamic Range 24 bit, 44.1khz digital FLAC version

Both LPs sound good but the percussion is audibly compressed versus the best digital versions. Say what you want about smoothness but both LPs are the second best versions of those records and lower fidelity than the best digital versions.

 

Definitely this, the dac/amp combo I am using is rated 10mW at 300ohm.

They are inherently different, 16 bit white noise at -96db with audacity will just give you 1 bit noise with 3 values, and it won't even reach -96dBFS, which is what should happen, dead silent. It sounded softer because there is a lot of 0s between +1s and -1s, unlike 24bit white noise at -96dB which has 8/9 bit worth of information. And the quantisation noise is definitely not the 1bit type as it is the difference between 24 and 16, it is kind of like you output the difference file in 16bit and will get nothing. The less difference heard from the 1bit noise is also because of whole bunch of 0 amplitude sample in the noise which doesn't affect the original sample at all.

btw I tried hiding 1kHz sine wave down -96dB in the music and downsample it to 16bit then compare them, the difference is still white noise, which makes sense I guess.

 

Meant -90db, but yes, makes sense. This is essentially what I meant but failed to express ...
It's interesting that changing the music by what's essentially 1 sample at a time at 16 bits is audible, however it has to be if the 16bit quantization noise can be audible.

Yes, right. Now that I think about it this is obvious ... It would always round to the best value, keeping the quantization noise less than 0.5LSB which means that a 16bit file can't capture the difference.
I somehow got confused somewhere along the way and then started confusing more things ... and it all ended up as a mess. Thanks for clearing that up.

 

Any interest in adding a DACCord (non-FF version) to the chart? I can loan one for testing, if interested.

 

This question has probably been asked many times before, but...

What's the difference between TCM and FRM filter modes in the Geek Out units? How does that translate to differences in sound sig?

 

FRM could be frequency response masking but no clue on the other and would not understand the difference if I did.

 

I've heard a lot of people say that with D/S DACs, the specific DAC chip doesn't matter so much, it's all implementation (which I assume means the power design, data interface and the analog section) - so I guess I'm a bit surprised that they can drop in a new DAC chip (4490) and immediately improve things. How is this done, or in other words, what are they improving? Is it primarily the filter implementations?

 

Sometimes I wonder if we are endeavoring to construct a mnemonic memory circuit using stone-knives and bear-skins.

 

Speaking of which, have you ever listened to a tube based Delta Sigma DAC? I recently found out there are a handfull that exist. Impressions I got from them is that the tubes add a warmth to counter the normally digital sound.

 

speaking of akm stuff. i actually got annoyed and wonder why most of all pro audio went circuss logic when i finally got back into audio. lucky i kept my old echo audiofire 2 cause echo no longer produces the 2/4 at all and the current echos all use circuss logic and, supposedly the ak4620 dac in my echo 2 is not bad in itself i found out. same happen to all other pro/studio stuff. so now have to pick between circuss logic dacs or delta-sigma dacs nowadays....