Can you tell 8-bit file from 16-bit file in a DBT? Listening test, RLJ

[pma]

I suspect that a lot of the difference is masked by the fact that the basic recording here is, frankly, a bit ******.

I agree, but you can tell the same about almost every recording of popular music and I do not think that many readers here listen much to classical music. Even our host does not and uses "female voices" for listening tests.

I may prepare the same test with classical music sample, but I assume that the result would be a bit different only in case if I choose a sample with very high dynamic range with long passages of ppp.

 

[pjug]

What do you mean by "appropriate dither?' Noise shaped?

View attachment 223713

To be able to tell the difference then requires people with acuity above 14 kHz.

Yes, I can't hear the shaped dither at normal speed, but playing at half speed you hear a lot of high pitch noise in the 8 bit track. I wonder how it sounds at 1X speed to someone who hears up to 18KHz or so.

 

[tmtomh]

I didn't ABX it, but the increased noise floor seems obvious during the opening acoustic guitar bit, as others have noted. Regardless, the 8-bit version with noise-shaped dither sounds very close to the 16-bit version. I do think the small differences are important and therefore the 16-bit sounds better in a way that is meaningful - but it does make me wonder how many bits it would take for me to not be able to notice any difference at all. I suspect only 10-12 bits, with proper noise-shaped dither, would start to sound indistinguishable for most of us.

It makes sense, too, as a first-rate professional reel-to-reel analogue tape deck reportedly is capable of an SNR equivalent to about 13 bits. No ability to shape the noise (aka tape hiss) there of course, but still, there are many stellar professional recordings out there that were made with such equipment and the noise floor on them is not noticeable, even if it is detectable.

 

[Steve Dallas]

The samples would be better if they had more high frequency information. Also, if it was a good song instead of an annoying one.

My Results...

Listening on speakers in my home office, I could not reliably hear any differences.

Switching to headphones, I could hear the elevated noise floor in the first several seconds and got it right 16/16.

Going back to speakers and cranking it up, I chose correctly 14 out of 16 times, having found what to listen for and discovered high SPL is needed to bring out the noise.

There is also a slight difference in the highest overtones of the guitar under critical listening with headphones but not with speakers in a room.

My main takeaway is how much impact masking has while listening to speakers in a room.

 

[Madlop26]

I Could not tell difference with my Revel speakers, but I could not crank up the volume, will try later with headphones or wait nobody is home. amazing a 8 bit can sound that good.

 

[Tangband]

When I do 2- channel recordings in 96 KHz / 24 bits, I always use dither in the DAW when I later in the process want to convert the file to 16 bit 44,1 KHz or when I do normalisation to -1 dB for the final mastering.

The noise from the microphones are at -85 dB so this dithering might probably be unnecessary ?

 

[ShadowNeko]

I agree with you, the difference in noise is clearly audible when listening to first 5 seconds only with high volume setting, that would be too much for the rest of the samples. Average rms power there is only -30 dBFS, so the noise compromise is audible.

foo_abx 2.0.2 report
foobar2000 v1.4.8
2022-08-12 09:54:15

File A: Rickie1.wav
SHA1: 1af7dad9413342c4c0eddab6182474bded3b02f0
File B: Rickie2.wav
SHA1: f23aede1fe13e20674ffde0e7e2f4d31880679c0

Output:
ASIO : Topping USB Audio Device
Crossfading: NO

09:54:15 : Test started.
09:54:43 : 01/01
09:54:49 : 02/02
09:54:55 : 03/03
09:55:08 : 04/04
09:55:16 : 05/05
09:55:22 : 06/06
09:55:34 : 07/07
09:55:47 : 08/08
09:55:53 : 09/09
09:56:08 : 10/10
09:56:17 : 11/11
09:56:26 : 12/12
09:56:38 : 13/13
09:56:47 : 14/14
09:56:56 : 15/15
09:57:05 : 16/16
09:57:05 : Test finished.

 ----------
Total: 16/16
Probability that you were guessing: 0.0%

 -- signature --
8b9e414650fbd6f914af71bb8b75b9b7ce16775f

It would be much more difficult if you tried more loud parts behind the first 6 seconds.

However, it is amazing what we get from 8-bit noise shaped file.

I tried again with the starting play point after the vocals start and i still passed 16/16 with the volume around 55-65db. It has this tinnitus sound during the whole track. I can hear till around 20khz.

 

[DonR]

I tried again with the starting play point after the vocals start and i still passed 16/16 with the volume around 55-65db. It has this tinnitus sound during the whole track. I can hear till around 20khz.

ahhh... I wondered why I heard no difference... damn tinnitus.

 

[amirm]

I agree, but you can tell the same about almost every recording of popular music and I do not think that many readers here listen much to classical music. Even our host does not and uses "female voices" for listening tests.

Don't make up nonsense like this. I listen to plenty of classical music. My last null test was vivaldi concert. Don't confuse what makes a good test track for some devices with what people listen to.

The track you used here has a ton of background noise which makes it a poor choice for such tests. Unless you have an agenda which you clearly have.

 

[danadam]

8-bit resolution version was prepared this way:

1) the original 44.1/16 file was upsampled using noise shaping dither, to 96kHz/8bit
2) the 96/8 file was then downsampled to 44.1/16 file

Upsampling with noise shaping is crucial here to keep acceptable low audibility of background noise.

Um... I don't think that's correct way. With 8 bit at 96k the noise is spread through wider bandwidth. When you convert that to 16 bit at 44.1k, that lower noise floor remains and it is not something, that would fit an actual 8 bit at 44.1k.

For example, here I started with flat dithered silence at 16 bit and did your procedure with 2x, 4x and 8x upsampling, only without noise shaping. The more you upsample in the first step, the lower noise floor you end up in the second one.

 

[RayDunzl]

This demonstrates how well dither actually works , if an 8bit file can be this good with dither , how good are then 16 and 24 bit .

1kHz tone, 16 and 24 bit, with and without dither.

Orange - 16 bit no dither
Green - 24 bit no dither
Black - 16 bit dithered
Blue - 24 bit dithered

 

[pma]

Um... I don't think that's correct way. With 8 bit at 96k the noise is spread through wider bandwidth. When you convert that to 16 bit at 44.1k, that lower noise floor remains and it is not something, that would fit an actual 8 bit at 44.1k.

That was of course intention to push the noise energy upwards, to make it less audible for human hearing. Please take into account that downsampling to 44.1/16 does not increase the resolution that was lost by upsampling to 96/8. If you do not like the method, you may compare the 96/8 byproduct with the 44.1/16 original, with the same result, or with original resampled to 96/16, again with the same result. The trouble with 8 bit files is that they are unsigned wav and not every software would play it, some will refuse to play. So I did not post the 96/8 vs. 96/16.
The key here is that the resolution in the manipulated file has remained 8 bit, even if the noise audible to humans was reduced. The noise energy, however, was not reduced, SNR as a ratio to rms noise remains only at 40 dB, as was shown in post #19. You get better SNR at lower frequencies at the expense of worse SNR at high frequencies, intentionally, as you know how the human hearing works.

@RayDunzl , when doing experiments with dither, please keep in mind there is not just one dither, but there are many versions - rectangular, TPD (both frequency independent) and then there is also a set of frequency dependent dithering methods - noise shaped that push the noise energy to high frequencies.

Noise shaping is also the method that almost all DACs work, if you feed them with 44.1/16 file they will internally upsample it to 384/32 or 768/32 format with noise shaping. The resolution still remains 16 dB, few dB of SNR is gained (shown in God of SINAD thread).

God of SINAD vs. reality we get from most available music files

God of SINAD vs. reality we get from most available music files There is a good thread here in ASR discussing the practical importance of SINAD number https://www.audiosciencereview.com/forum/index.php?threads/could-we-all-be-wrong-about-sinad.26480/ which links to the original article at...

www.audiosciencereview.com

 

[pma]

I understand that the process is not straightforward to everyone. For this reason and for those who may have doubts, I am posting the true 96kHz/8bit file that was downsampled to 44.1/16 for the test in post #1.

https://pmacura.cz/Rickie2_96-8.zip

However, I cannot guarantee that it will be playable with every SW and every DAC. In my case, it works with Foobar and Topping D10s DAC.

and the histogram

Yes we have no more than 8 bit resolution.

 

[danadam]

Please take into account that downsampling to 44.1/16 does not increase the resolution that was lost by upsampling to 96/8.

True, but the resolution of 96/8 (in the audible range) is not the same as of 44/8. Every doubling of sampling frequency lowers the noise floor in the audible range by 3 dB (with flat dither, with noise shaping even more). Listen for yourself (flat dither used):

• 1_input.flac - -24 dBFS 1k sin, generated at 44/16
• 2_actual_8bit.flac - the input converted to 44/8 and then to 44/16
• 3_not_really_8bit.flac - the input converted to 352/8 and then to 44/16

1 and 2 is the correct 8 vs 16 bit comparison, 1 and 3 is not.

The trouble with 8 bit files is that they are unsigned wav and not every software would play it, some will refuse to play.

So convert this 96/8 to 44/8 and then the 44/8 to 44/16 and post that.

Here's how 44/8 really sounds, 5 second snippets:

• Rickie3.flac - flat dither
• Rickie4.flac - dither with noise shaping

You can even check that the 8 least significant bits are set to 0:

]$ hexdump -s$((44 + 2*2*40848)) -v -e '1/2 "%04x\n"' Rickie3.wav | head -n 10
0600
1400
0700
1600
0700
1700
0800
1a00
0800
1b00
]$ hexdump -s$((44 + 2*2*40848)) -v -e '1/2 "%04x\n"' Rickie4.wav | head -n 10
0e00
1500
0100
1100
0900
2100
0a00
0a00
0200
2e00

 

[danadam]

That was of course intention to push the noise energy upwards, to make it less audible for human hearing.

But you pushed it beyond the limits of 44/8 container. If this was the intention, then I'm not sure what's the point of this test. Everyone knows that with enough upsampling even 1 bit sounds good. After all DSD exists.

Noise shaping is also the method that almost all DACs work, if you feed them with 44.1/16 file they will internally upsample it to 384/32 or 768/32 format with noise shaping.

Really? What noise is there to shape when you increase bitdepth?

 

[pma]

Illustration of upsampling effect on 1kHz sine spectrum

Original file is 1kHz sine in 44.1kHz/16bit format. This was upsampled to 96kHz/8bit format noise shaped (white) and to 96kHz/16bit format (blue). The resulting spectra were compared in DeltaWave software

DeltaWave Audio Null Comparator | DeltaWave documentation

deltaw.org

For the purpose of the test (post #1) the original file (blue) was upsampled to 96kHz/8bit format noise shaped and then downsampled to 44.1kHz/16bit format to keep the same format for the purpose of the ABX listening test (to prevent easy identification from DAC display once showing 44.1 and then 96). This only cuts unnecessary frequencies above 22kHz. Otherwise the spectra are identical below 22kHz with the 96kHz files.

With upsampling, the quantization noise is spread to wider frequency band, DC – K*Fs/2. When it is noise shaped, the noise energy is pushed to high frequencies and we get better SNR at low frequencies.

Comparing the 16bit original with the 8bit upsampled noise shaped spectrum, we can see that the effective number of bits (ENOB) differs considerably with the bandwidth of interest

ENOB = (SINAD – 1.76dB)/6.02 or sometimes (SNR - 1.76dB)/6.02

ENOB is almost same for the 16bit original and 8bit noise shaped up to 200Hz. Between 200Hz and 10kHz, ENOB of the 8bit noise shaped signal is still acceptable. That is why we do not hear much noise in the 8bit test file (Rickie1). Above 10kHz, ENOB of the 8bit noise shaped signal falls very quickly resulting in a poor SINAD = 40dB when defined over the whole 22kHz band without frequency weighting. This is the psychoacoustics in practice.

 

[tmtomh]

That was of course intention to push the noise energy upwards, to make it less audible for human hearing. Please take into account that downsampling to 44.1/16 does not increase the resolution that was lost by upsampling to 96/8. If you do not like the method, you may compare the 96/8 byproduct with the 44.1/16 original, with the same result, or with original resampled to 96/16, again with the same result.

I think the point @danadam is trying to make (and danadam, please correct me if I'm mistaken) is why not compare a 44.1/8-bit byproduct with the 44.1/16-bit original? That's the way to isolate the variable of bit-depth. If I read danadam's comments correctly, then by increasing the sample rate, you've effectively rigged the comparison in favor of the loss of 8 bits being less audible than it would otherwise be.

 

[maty]

Very easy test. With JRiver, without ABX, you know.

 

[danadam]

I think the point @danadam is trying to make (and danadam, please correct me if I'm mistaken) is why not compare a 44.1/8-bit byproduct with the 44.1/16-bit original? That's the way to isolate the variable of bit-depth. If I read danadam's comments correctly, then by increasing the sample rate, you've effectively rigged the comparison in favor of the loss of 8 bits being less audible than it would otherwise be.

Yes, basically that.

As it is right now, it rather creates confusion than illustrates anything. The title says "Can you tell 8-bit file from 16-bit", the included files are both 44k, so it is natural to assume, that this is "8-bit @ 44k vs 16-bit @ 44k", but effectively it is "8-bit @ 96k vs 16-bit @ 44k". Those are two different things.

Or put it another way, as I already mentioned DSD. What would you think (general you) if I took 16/44k file, converted it to DSD, then back to 16/44k and posted those two 16/44k files in a thread titled "Can you tell 1-bit file from 16-bit"? Well, most would probably figured out that DSD was involved, but overall it would be rather clickbaity/disingenuous.

 

[jlo]

The concept of dither is so obscure that many in music production don't understand it either.

To see and listen to the effect of dither and quantification, I commited this free software : https://www.ohl.to/about-audio/audio-softwares/ditherer
It was a long time ago but it should work on any MS windows version (or as a VST on other OS)