Skip to main content

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: 16 bit vs 24 bit (Read 253834 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

16 bit vs 24 bit

Reply #100
~26,000 views... and still a mystery.  ;>

16 bit vs 24 bit

Reply #101
18/32 - there's a harshness in the left channel around the 8.5 mark, which I had originally mistaken for ground hum, which I felt I could possibly distinguish in character between 16hpt and 24. I started pretty good at 5/6 but I felt like I could not hold on to the difference.

AndyH, I took the above to mean that Axon initially used an apparent difference 8.5 seconds into the track as a basis for entering his ABX responses; and that he got 5 out of 6 correct for the first 6 entered responses relying on that technique.  However his concentration/hearing then began to fail him and he ended up with only 18 out of 32 correct after giving the 32nd response.

This is reminiscent of my hearing fatigue issue.  But, in my case, if I find I can no longer hear a difference I stop and wait for my hearing/concentration to recover.

I note we are still waiting for Martin to tell us the subjective difference he heard when he was doing his 100 listening trials, with 62 correct answers.  That's assuming the difference had a definite "quality" he perceived.

I myself find music at 24 bits sound more "liquid" than the same music file truncated to 16 bits.  But if it is properly dithered down to 16 bits, I cannot perceive a difference when 24-bit music is changed to 16 bits.

Has anyone else heard a difference at the 8.5 second mark?  Does this difference disappear if another form of dither is used? (I don't have access to playback equipment at the moment to check out the sound at the 8.5 second mark myself.)

~26,000 views... and still a mystery.  ;>

It seems no-one can actually supply a music sample (recorded at a normal rather than very quiet level) that demonstrably sounds different when dithered down to 16 bits.  Martin has a music sample that might possibly qualify, if his own dithering and ABXing are to be accepted at face value, but he has not responded to AndyH's query regarding what the difference sounded like. And there is a dearth of ABX submissions from other HydrogenAudio members, though Axon as made a verbal report.  I can only presume a few people listened to Martin's samples and heard no difference, but have not bothered to report this.  I personally regard the technical issue as important, as many Bluray discs are being released with lossless tracks at a 16-bit depth (dithered), rather than a 24-bit version of the master.  Some people are prepared to pay a premium for a Blu-ray disk that provides 24-bit sound.  They may actually shun a Bluray that only has 16-bit depth audio.  Are such people suffering under a misapprehension?

16 bit vs 24 bit

Reply #102
Hi guys,

sorry for the late reply, been very busy for a while with a Critical Listening class...

Anyway, you wanted to know what I heard in David's dithered 16-bit version. Once again I found myself focusing on the tone of the brass, the 16 bit was thinner, had less body or less punch if you like. Remember that I was listening on my crappy laptop setup with very low level on the phones, so had to listen to changes in timbre and not in low level detail or similar. As a matter of fact, I tried turning down the level even more and it possibly made it even easier to hear which one was which. If you're interested, with the other "special" dither I didn't hear this effect, but another one instead which seemed to affect the stereo perspective. 

Have no explanation to why I hear this. It does not make any sense, but that's exactly why I find it an interesting subject worth studying further. If lower bit depths have an effect on the musical tone itself (and not some tiny technical loss of quality) we should perhaps be worried. Especially if we're stuck with dithered 16 bit as the best delivery format ever.

Mentioned earlier that several factors are important when listening. To me, these samples work well but the equipment was far from optimal. If you can't hear a difference easily, it might also have something to do with my ears, I work with audio full time and have done so for the last twenty years, and I believe that our hearing can improve with training. If you try listening, a straight signal path without any signal processing is essential, my noise canceling headphones were useless.

Martin

16 bit vs 24 bit

Reply #103
Anyway, you wanted to know what I heard in David's dithered 16-bit version...
Just to clarify, "my" version, here...
http://www.hydrogenaudio.org/forums/index....mp;#entry610883
...is "Your 16 bit file, converted back to 24-bits, with a little high-pass noise added to make sure the bottom 8 bits are moving".

This isn't "my" attempt to dither it (IMO there's nothing wrong with your dithering, but no one has provided an alternative rendition here) - it's simply my attempt to remove 16bit vs 24bit sound card peculiarities by filling the bottom 8 bits with something relatively benign.

What I provided is further away from the original than even your 16-bit version - as such, it should be easier to hear a difference, unless the difference is down to the sound card's handling of 16-bit vs 24-bit material.


I have to admit, I've even had a little Google this morning to check there's no organised hoax against HA - we've got Martin Kantola ABXing the inaudible in this thread, and BORK ABXing the inaudible in the lossyWAV thread.

If other people can confirm, it's quite exciting, because it takes ABXing to another level.

Cheers,
David.

16 bit vs 24 bit

Reply #104
P.S. I ran the track 24.wav through the latest version of lossyWAV (1.1.2d).

In insane mode, it quite often wanted to keep 19 bits.

In standard mode, it quite often wanted to keep 17-18 bits.

(I'm judging this by the amount of noise it adds, not by reading the bits_to_remove values directly, so it's not 100% reliable).

Cheers,
David.

16 bit vs 24 bit

Reply #105
David’s first remarks about his lossyWAV results made me curious about what I could find by a different route. The results were interesting, in an academic sort of way, but I’m not certain how to interpret them. To report seemed like a lot of writing, maybe with no purpose, so I moved on to other projects. Then  David’s second post on the subject decide me to try to write up my experiment and seek comment.

Since this concept is important to understand in order to follow my process and results, I first describe the following, just in case there are any readers not already familiar with it.

If one takes any on-computer audio file, it is easy to make an exact copy. If one then inverts one of the two identical files (easy done in most any audio editor) and mixes it with the other (not inverted), aligning sample for sample, the two completely cancel each other. The result is silence, digital zeros.

If one takes any two different audio files and mixes them together, one has a combination, such as a vocal and backing track now in one file. If one then takes one of the two, inverts it, and mixes it with the combination, as in the above paragraph, one is left with the original of the other file -- almost.

In the first mixing, each resulting sample is the sum of the corresponding samples from the two inputs. A quantization error occurs in every case where the sum of the two exceeds the precision the format can carry. This will occur even when working in 32 bit float.

So, if one then inverts and mixes in the original of the second file to cancel it, one does not get digital zeros, one gets the quantization error values. In a sample I tried, working in 32 bit float, the result had a peak amplitude of -144dB and an average RMS amplitude of -155dB. Silence in any practical sense, but not complete.

To start the following, I opened the provided 24 bit, 48kHz file into 32 bit float, the normal format for CoolEdit. This changes no values, it only provides greater working precision. In all cases, unless otherwise specified, a 16 bit result was converted back to 32 bit in order to compare it against any 32 bit originals. Again, re-converting to 32 bit changes no values that were in the 16 bit version, everything below 16 bits is just zeros.

A word about dither. Comments are welcome. CoolEdit provides three parameters for dither. I don’t recall that any others are possible. One is Dither Depth, the amount of dither. Another is p.d.f. or dither type (such as rectangular, triangular, etc.). The third is noise shaping. My program version has 12 choices for noise shaping.

Dither depth is the most important of these parameters. Too little and quantization distortion still exists. Too much and unnecessary noise is present in the final result. The dither type effects the results, such as the potential for modulation, and noise shaping is mainly important for distributing the added noise for least audible impact.

As long as the dither depth is good, one can leave out noise shaping, and use any p.d.f., to get very good, if not necessarily optimum, results. While there are many favorite dithering variations, they all, I believe, come down to very difficult to notice differences that are unimportant in normal listening.

dithered silence
If one generates a file of silence, digital zeros, in float format, one can then convert that to 16 bit. The result will still be silence; there are no values except zero, so no quantization errors occur.

One can apply dither to the 32 bit silence and then convert to 16 bits. The resultant will contain only the noise that gets added to any music file one converts in the same way. This consists of the dither noise and the quantization noise from converting it to 16 bit.

Once upon a time, by experimenting with various combinations, I selected the dither settings that I use with all my music. It results in the least amount of added noise (that works) and is completely inaudible to me on my equipment. The “that works” part means that less dither does not eliminate all the quantization distortion. This is very hard to determine with music, but fairly easy using pure test tones.
Dither depth 0.5 bits, p.d.f. Shaped Triangular, noise shaping C1

Now finally, the real gist of the report. First I converted the provide sample of 24/48 (now 32/48)  music to 16 bit using my normal dither, then converted it back to 32 bit. This, you recall, leaves everything below 16 bits as zeros. I then inverted and mixed it with the original to provide a difference.

I realize that this process is considered useless with perceptual encodings. The good reason for this view is that the important aspects of perceptual encodings are based on human hearing, not on any data differences. Data differences might be large but irrelevant. In this case, while I am not sure about how to interpret the results (thus this lengthy report, for critique and comment), I am not sure it is so reasonable to lump them with the perceptual encoding viewpoint.

dithered conversion
The result of these three steps is essentially dithered silence, as described above. With the tools CoolEdit provides, I cannot tell any difference between this and the dithered silence results in Spectral View, with the Frequency Analysis graph, or by listening. There is some difference in Statistics (peak amplitude), but I can’t say if it is due to anything other than quantization noise. The columns are left and right channels.
                            dithered silence                  music
Peak Amplitude:        -78.27 dB       -78.27 dB       -75.42 dB       -75.55 dB
Minimum RMS :          -88.42 dB       -88.45 dB       -89.13 dB       -89.14 dB
Maximum RMS :        -87.24 dB       -87.29 dB       -87.36 dB       -87.45 dB
Average RMS :          -87.87 dB       -87.86 dB       -88.22 dB       -88.22 dB
Total RMS :              -87.87 dB       -87.85 dB       -88.22 dB       -88.21 dB

non-dithered conversion
Repeating the procedure, but without dithering, gives these lower statistics. Spectral View show a very even noise distribution, except a slightly greater intensity at very low frequencies. Frequency Analysis show a linear rise of 20dB below 140Hz to the lower limit of frequency resolution, otherwise a flat line over the audio spectrum. Although at a considerably lower level than the dithered version final result, this is quite audible-- at the highest output of my headphone amplifier.
Peak Amplitude:      -96.36 dB   -96.36 dB
Minimum RMS :      -101.49 dB   -101.42 dB
Maximum RMS :      -100.84 dB   -100.83 dB
Average RMS :      -101.11 dB   -101.11 dB
Total RMS :      -101.11 dB   -101.11 dB

Since CoolEdit provides no control point between applying dither and converting to 16 bits, I created my own dither file by generating a file of low intensity white noise in 32 bit float. Its statistics are
Peak Amplitude:      -80.14 dB   -80.14 dB
Minimum RMS :      -89.88 dB   -89.88 dB
Maximum RMS :      -89.09 dB   -89.09 dB
Average RMS :      -89.49 dB   -89.49 dB
Total RMS :         -89.49 dB   -89.49 dB
Its Frequency Analysis graph is a flat line across the audio spectrum.

As proof of concept, I tested with a created file consisting of 10 seconds of a 3kHz tone followed by 10 seconds of digital silence. For the 3kHz part:
            Mono
Peak Amplitude:      -78 dB
Minimum RMS :      -81.02 dB
Maximum RMS :      -81.02 dB
Average RMS :      -81.02 dB
Total RMS :      -81.02 dB
The 3kHz tone was chosen to be easy to distinguish the result of dithering vs non-dithering. Non-dithered conversion to 16 bit produces ample distortion, easily viewed and measured. Then, starting over,

(1) Mix same 32 bit audio file with 32 bit white noise file, producing dithered audio.
(2) Convert mixed file to 16 bit, no dithering. Produces non-distorted 16 bit.
(3) Convert back to 32 bit. Nothing below 16 bit but zeros.
(4) Convert white noise file 16 bit, no dithering.
(5) Convert 16 bit white noise back to 32 bit. As always, this leave zeros below 16 bits.
(6) Invert (5) and mix with (3).

The second half of the resulting file, which was originally silence, is again complete silence. The first half of the file, the tone, is now the tone plus quantization noise.

(7) Invert the original file and mix with (6). The second half of the file is still zeros, the first half is only the quantization noise, and whatever was below 16 bits in the original tone. This is seen to be essentially flat across the audio spectrum, with a slight rise at the lowest frequencies.
Peak Amplitude:      -90.31 dB
Minimum RMS :      -98.97 dB
Maximum RMS :      -98.23 dB
Average RMS :      -98.58 dB
Total RMS :      -98.58 dB

Having demonstrated the feasibility, I did the same thing with the provided audio sample (now in 32 bit format for convenience). Call the original audio A and the generated white noise B.

(1) Mix A and B
(2) Convert mixture (1) to 16 bit, no dithering.
(3) Convert (2) to 32 bit.
(4) Convert B to 16 bit, no dithering.
(5) Convert (4) to 32 bit.
(6) Invert (5).
(7) Mix (5) with (3), eliminating dither of mixed A + B.
(8) Invert (7).
(9) Mix (7) with A, leaving only the quantization noise and whatever was below 16 bits in A.

The final result:
Peak Amplitude:   -90.35 dB   -90.35 dB
Minimum RMS:   -98.49 dB   -98.53 dB
Maximum RMS:   -97.71 dB   -97.67 dB
Average RMS:   -98.1 dB   -98.1 dB
Total RMS:   -98.1 dB   -98.1 dB

This is whatever was below 16 bit in the original A plus quantization error noise. I believe the levels are higher than the non-dithered conversion reported above only because it includes additional quantization noise from the white noise dither. The spectra is flat, the sound is everywhere a light hiss, audible only at a very high volume setting. If there is any real difference between the 24 bit and 16 bit versions, aside from the dither and quantization noise, it is below the resolution of my tools. I suggest this is indeed, as David commented, inaudible. Any ideas or suggestions?

Unfortunately, if looks like my attempt to format so that the numbers display in aligned columns is not working. Also, this new software, with the entry window so much wider than the screen, requiring scrolling right and left, is a real pain.

16 bit vs 24 bit

Reply #106
Nulling does unfortunately not reveal everything, since we actually don't know what we are looking at or listening to. What is noise? Don't want to repeat myself so please have a look at my post in this thread in Gearslutz

Martin

16 bit vs 24 bit

Reply #107
I'm posting in this thread because I find it fascinating and want to learn, not to criticise anyone. I wanted to say that straight away, in case it sounds like I do criticise anyone!

AndyH-ha, I followed all of that. I concur with almost all of it. I know we've both been using CEP professionally for a long time now - though for quite different uses. I only use it as a "toy" in the field where you use it professionally. I have, but don't use, Adobe Audition. That's probably irrelevant to the current discussion.


I'm not sure about this bit...
Quote
If one takes any two different audio files and mixes them together, one has a combination, such as a vocal and backing track now in one file. If one then takes one of the two, inverts it, and mixes it with the combination, as in the above paragraph, one is left with the original of the other file -- almost.

In the first mixing, each resulting sample is the sum of the corresponding samples from the two inputs. A quantization error occurs in every case where the sum of the two exceeds the precision the format can carry. This will occur even when working in 32 bit float.
If you don't scale and don't clip - and the originals do not clip - then you should still get digital silence. It certainly works here (CEP 1.2a!). It won't work if you reduce each file to 50% during the addition (because this implies scaling and re-quantisation), but it will work if neither file peaks above 50% to start with, and you add them as-is (no scaling, no clipping, no actual re-quantisation because you're just adding two integers). It works at 16-bit or 32-bit.

If the original 16-bit audio hits negative full scale, then you can't invert it properly, because negative full scale is one sample larger than positive full scale (-32768 vs +32767). CEP reduces these clipped negative samples by 1 LSB when transforming them to full scale positive samples using the function "invert", but keeps them intact if you tick the "invert" tick-box in the mix-paste dialogue.


I think the increase at low frequencies you've seen in some experiments could be a quirk of CEP's rounding - it certainly rounds upwards in some operations, thus introducing a maximum of one LSB (average of 0.5 LSB) DC shift. I think the FFT is showing this. The exact appearance depends on the FFT window, but what you'd expect to simply fall into the "DC" bin (and so not be shown on that graph, if it's plotted correctly) ends up being distributed across many of the low frequency bins. In transform>amplitude>amplify you can add an intentional DC shift - add it to digital silence, and look at the frequency analysis - strange, isn't it?


Then we get to the crux of the problem. We all know the mathematical difference between the 16-bit and 24-bit files is almost nothing. It's a struggle to hear it without the music in top, never mind with!

Martin suggests this isn't the point, because hearing doesn't work like that, and then gives two (IMO unconvincing) examples...

Quote
First, I truncated (no dither) the file to only 8 bits. Then I performed a null test, comparing it to the 24 bit version. As Ethan and others have observed, mostly noise and definitely no music there.

BUT, when I took this 'noise' (the result of the null test) and combined it with the truncated 8 bit version, it really helped to 'restore' the music. So, the conclusion would be that we can't simply listen to a null test difference and say it's only noise and not music related information.
I don't know how your ears work, but to my ears, the 24-bit file sounds fine, the 8-bit file sound horribly noisy, and the "difference" / "null" file sounds exactly like the noise in the 8-bit file, but without the music. I don't hear any magic - there's "music", "music + noise", and "noise". In my mind, if you remove the "noise" from the "music + noise" I'd expect it to sound like "music". It doesn't surprise me at all.

Also, elsewhere in that thread, there's theexample of a low pass filter - again suggesting that if you listen to the "dull" low pass filtered version, and then listen to only what you've removed (a hissy, tinny thing), you wouldn't dream that putting the two together would give you the nice bright full sounding original. Well I would - that's exactly what my ears tell me to expect!


So, I think they're really bad examples. They don't work for my ears at all.
Also, FWIW, and IMO, the logic of some members on the other board, and the understanding of "resolution" especially, is really poor. They kind of got there in the end - but what's being discussed is fundamental and important, while the understanding is really lacking.

However - there are situations where the mathematical difference between two signals doesn't related to the audible difference between those two signals. As anyone who has ever done .mp3 minus .wav can tell you, the mathematical difference is huge, the audible difference can be tiny. Listen to the mathematical difference ("null") all you like - you'll never hear it fully in the mp3 itself as the difference between the mp3 and the wav. Here, my ears tell me that "music" (original .wav), "music + noise" (.mp3), and "noise" (.mp3 minus .wav) cannot possibly be related in the way that I know (mathematically) they are.

So, it's possible that the audible difference between a 16-bit file and a 24-file is different from the mathematical difference. Surprising, but possible.

The classic paper(s) from Lipshitz and Vanderkooy, e.g. http://www.aes.org/e-lib/browse.cfm?elib=11586 (I think - my copy is on paper somewhere!) point out that while the mathematical difference (due to correctly dithered truncation) "sounds" like pure uncorrelated noise, if you examine it properly there is still third order and higher correlation present - it's just the first two orders of correlation (which give rise to harmonic distortion, and noise modulation) have been removed. This is much better than not dithering, but not perfect.

So, can we hear those third order and higher errors? Rather than simply hear the added noise by cranking the volume control up ridiculously high? As far as I know, this is the first time anyone has ABXed this, so it's well worth looking at.
If the equipment is working properly, at a reasonable listening level, and digital silence vs dithered silence cannot be ABXed, and 24-bit vs dithered 16-bit can be ABXed, then that is exciting.

I don't know if this is what we have here. lossyWAV thinks the dither might be audible anyway, so how loud is Martin playing these files?

Martin: can you ABX the "null" file vs digital silence? At the same listening loudness that you can ABX 24-bits vs 16-bits?

Cheers,
David.

16 bit vs 24 bit

Reply #108
Greetings, found your forum recently, so here's my first post! Wanted to share another set of files that might be useful for listening tests. The source material is a 'raw' unprocessed master recording downsampled from 96kHz. Uploaded the files here:

Digital audio resolution test files

Below you can see a quick ABX result I got using my laptop and a pair of Beyerdynamic DT250 headphones.



The second set of files are from an old master tape, comparing 24 bit versus 12 bit versus 320kbps mp3.


I would sincerily hope that *nobody* would try to do *anything* definitiive with these files. Based on my evaluaton of the 24/96 version, here's not enough energy > 20 Khz (or even 5 Khz!) to be meaningful, and the  dynamic range is only about 60 dB.  If you want to convince yourself that 12 KHz sampling and 12 bits are pretty good, be my guest!

16 bit vs 24 bit

Reply #109
I know I can convert properly; possible audible differences with other conversions to 16 bit are of no value. Tests based on them on very suspect. I understand most anyone can do a proper conversion, given the necessary software, but a fuller specification of the particular process employed is desirable.


What exactly do you need to know apart from the dither used? Please let me know. While I understand that this is slightly controversial, we have to remember that most professional audio is done in 24 bit today, probably for good reason. So we could at least suspect an audible difference. But it's a good idea to download only the source file to save download time.

Blue printing on a black background is extremely difficult to read.


Oh, I'm sorry, looked fine on my screen, but fixed it now! Please reload the page. Thanks for pointing that out.

Martin




It would be helpful if Martin Kantola could describe the differences he heard, where in the file he heard them, and most importantly can verify that his laptop plays back 16-bit and 24-bit digital audio correctly.


While the hardware of my laptop is 24bit/192kHz capable, the specs are not exactly impressive:



Will get back to the differences I heard, have to try a 16 bit padded to 24 bits to make sure it works as it should.

Martin


IME, The RealTek audio interfaces aren't really 24/192 capable. FR testing shows that they downsample everything to a much lower sample rate, and much lower resolution.



16 bit vs 24 bit

Reply #110
I would sincerily hope that *nobody* would try to do *anything* definitiive with these files. Based on my evaluaton of the 24/96 version...
There are two completely different recordings on that page. I don't think anyone in this thread has shown any interest in the second one.

Cheers,
David.

16 bit vs 24 bit

Reply #111
Martin suggests this isn't the point, because hearing doesn't work like that, and then gives two (IMO unconvincing) examples...

I don't know how your ears work, but to my ears, the 24-bit file sounds fine, the 8-bit file sound horribly noisy, and the "difference" / "null" file sounds exactly like the noise in the 8-bit file, but without the music. I don't hear any magic - there's "music", "music + noise", and "noise". In my mind, if you remove the "noise" from the "music + noise" I'd expect it to sound like "music". It doesn't surprise me at all.


Let me try to explain myself better, the "magic" would be in the last, third file, the "nullresult-plus-truncated.mp3". By adding only what sounds like noise to the truncated 8-bit file, two things happen, at least to my ears:

1. The noise goes away.
2. The distortion goes away.

What I tried to show was that we clearly removed more than "just noise" when truncating, and so the removed bits must contain signal-related information. Adding any other uncorrelated noise (with the same spectrum) to the truncated file would not do the trick. Yes, as such it's just an obvious subtraction-addition operation and shouldn't surprise anybody. But the point is that we can't necessarily tell by analyzing or listening to the null "noise" if there's useful information or not in there, and certainly not how much.

I don't know if this is what we have here. lossyWAV thinks the dither might be audible anyway, so how loud is Martin playing these files?


Not too loud at all, mainly because of the high impedance (250 Ohms) of the headphones. Can't measure SPL from headphones.

Martin: can you ABX the "null" file vs digital silence? At the same listening loudness that you can ABX 24-bits vs 16-bits?


Of course not, but that's a good point. As explained above, the null file in itself is meaningless to us. The LSB information only makes sense when combined with the rest of the data.

Martin

16 bit vs 24 bit

Reply #112
I would sincerily hope that *nobody* would try to do *anything* definitiive with these files. Based on my evaluaton of the 24/96 version, here's not enough energy > 20 Khz (or even 5 Khz!) to be meaningful, and the  dynamic range is only about 60 dB.  If you want to convince yourself that 12 KHz sampling and 12 bits are pretty good, be my guest!


To begin with, the old analog tape is a fantastic recording, and regardless of the tape noise I personally find it to be of very high quality. But thanks for bringing up the 12 kHz 12 bit suggestion. Am I correct to assume that you're saying this particular recording could be truthfully captured with said digital resolution based on the dynamic range and frequency range you estimated? If so, that's where I find the sampling theorem not properly applied. Why? Because it's not trivial to determine exactly where the music ends and the noise begins. We can guess of course, but in order to make sure we must capture the tape noise as part of the signal too. And so an interesting question forms, what is the noise floor of the tape noise?

Martin

16 bit vs 24 bit

Reply #113
What I tried to show was that we clearly removed more than "just noise" when truncating, and so the removed bits must contain signal-related information. Adding any other uncorrelated noise (with the same spectrum) to the truncated file would not do the trick.
I don't think you've proven the point (in the scientific sense).

Where you're making it look more clever than it is is in saying "removed". Forget that. Say "added" instead. And if you don't like that, remember than you can add a negative, so it means the same thing.

Then the situation becomes....

Code: [Select]
1 = signal

2 = signal + noise_A

3 = 2 - noise_A
  = (signal + noise_A) - noise_A
  = signal
  = 1

...where
1 = original 24-bit signal
2 = dithered 16-bit signal (or 8-bit signal in this case)
3 = difference ("null")

You believe there's something magical (or, specifically, signal related) about noise_A because, when you subtract it, you get the signal back - whereas if you subtracted "any other uncorrelated noise (with the same spectrum)" (let's call that noise_B) you would not. That's not magic at all - that's very simple maths.

Subtracting noise_B doesn't give you the original, because you didn't add noise_B in the first place. If you had done, it would. It wouldn't prove that noise_B was related to the signal either.


I don't dispute the overall point you're trying to make - noise_A might be related to the signal. In fact, it demonstrably is (if you know what to look for, you can create waveforms where this fact is clearly visible). But what you've done doesn't prove it. We need to look deeper than that.


Here's an interesting question (for the 16-bit version at least): Does any similar (but completely uncorrelated) noise make it sound worse, or is it only dithering+truncation that makes it sound worse? In other words, is noise_A a special "bad-case", or not?

Cheers,
David.

16 bit vs 24 bit

Reply #114
what is the noise floor of the tape noise?
5-bits.

(someone told me once - I have no better reference than that!).

i.e. if you have truly random white noise, and you want to sample it, you can do so transparently with 5-bits.

If it's not white or not truly random, then you might need more. (I could go on about those two for a page or more, but won't).


I can't remember how many bits lossyWAV allocates, but it's about right IIRC.

Cheers,
David.

16 bit vs 24 bit

Reply #115
On the dither_noise_test sample you supplied lossyWAV removes 10 bits [edit] at --portable; 9.6821 at --standard [/edit] - which is at [edit] (or very close to) [/edit] the hard internal limit of 5 bits remaining + sign bit.
lossyWAV -q X -a 4 -s h -A --feedback 2 --limit 15848 --scale 0.5 | FLAC -5 -e -p -b 512 -P=4096 -S- (having set foobar to output 24-bit PCM; scaling by 0.5 gives the ANS headroom to work)

16 bit vs 24 bit

Reply #116
BTW...
Code: [Select]
foo_abx 1.3.1 report
foobar2000 v0.9.4.2
2009/02/09 16:29:47

File A: D:\audio\digital audio resolution test files\96k24b.wav
File B: D:\audio\digital audio resolution test files\mp3-48k320kbps.wav

16:29:47 : Test started.
16:31:04 : 01/01  50.0%
16:31:25 : 02/02  25.0%
16:31:39 : 03/03  12.5%
16:31:50 : 04/04  6.3%
16:32:12 : 05/05  3.1%
16:32:20 : 06/06  1.6%
16:32:30 : 07/07  0.8%
16:32:43 : 08/08  0.4%
16:32:44 : Test finished.

----------
Total: 8/8 (0.4%)

Easy.

The "T" of "I'm travelling light" is changed by the mp3 encoding.

This is the first time I listened to it. I agree it's a nice recording in many ways. It would have to be butchered by noise reduction to go onto a CD though. Play that as-is on the radio today and the hiss would be dragged up to be louder than the music.

I don't know what mp3 encoder + settings you used (and hissy recordings are a challenge to mp3 in general) but I bet it can be encoded better than that.

Cheers,
David.

P.S. EDIT: That "24/96" recording looks like it was copied in the analogue domain from a CD. There's a null at 22kHz, a little crap (completely unrelated to the music) a little above it, and then nothing but noise, tens of dB down, above that. I hope Arny was joking somewhat with his figures though - there's clearly real audio information in the speech transients up to 20kHz, and a 6kHz LPF is easily detected. While I'm sure you could take the bitdepth right down (especially as it's a 96k recording), lossyWAV wants to keep 16 bits at a few points. The recording, as given, only peaks at -13dB, so the top 2 bits aren't being used either. Maybe it gets louder later on.

16 bit vs 24 bit

Reply #117
Total: 8/8 (0.4%)
Easy. The "T" of "I'm travelling light" is changed by the mp3 encoding.


Yep, that's how it worked for me too, find a little piece in the music that changes and focus on it. Thanks for listening!


This is the first time I listened to it. I agree it's a nice recording in many ways. It would have to be butchered by noise reduction to go onto a CD though. Play that as-is on the radio today and the hiss would be dragged up to be louder than the music.


I should tell you that this is only a fragment of the whole piece of music, it gets much louder as you guessed, so the overall dynamic range is better. To de-noise or not is an interesting question.


I don't know what mp3 encoder + settings you used (and hissy recordings are a challenge to mp3 in general) but I bet it can be encoded better than that.


The source is there, so if you want to give it a shot go ahead. I'm sure there are differences in encoders.

P.S. EDIT: That "24/96" recording looks like it was copied in the analogue domain from a CD. There's a null at 22kHz, a little crap (completely unrelated to the music) a little above it, and then nothing but noise, tens of dB down, above that.


Interesting, don't know the exact signal path that recording has been through. Wasn't born when it was made. :-)

Martin

16 bit vs 24 bit

Reply #118
Could this whole situation be simply from Martin's poor-quality audio card causing some kind of non-linear distortion that makes these results actually audible when normally they are not?

16 bit vs 24 bit

Reply #119
You believe there's something magical...


Oh, not at all, that's just a word you used and I liked it ;-)

I don't dispute the overall point you're trying to make - noise_A might be related to the signal. In fact, it demonstrably is (if you know what to look for, you can create waveforms where this fact is clearly visible).


Good, that's all I wanted to say.

But what you've done doesn't prove it. We need to look deeper than that.


Well, even if I wasn't out to prove anything scientifically, I do feel that my simple test shows that the null noise is more related to the music than not. Just try adding any uncorrelated noise instead. But I agree, we need to look deeper.

Here's an interesting question (for the 16-bit version at least): Does any similar (but completely uncorrelated) noise make it sound worse, or is it only dithering+truncation that makes it sound worse? In other words, is noise_A a special "bad-case", or not?


Not sure I follow you 100% here, please help me understand, but you make the assumption that noise_A is only... noise?

Martin

16 bit vs 24 bit

Reply #120
Could this whole situation be simply from Martin's poor-quality audio card causing some kind of non-linear distortion that makes these results actually audible when normally they are not?


The poor quality of the equipment is actually what makes this interesting. But we've discussed using 24 bit files padded with zeroes instead of 16 bit ones to rule out some possibilities of contamination.

As far as audible on a good system, I would go as far as claiming that any full time mastering engineer not able to hear the difference between the 24 bit source and the 16 bit on his own setup should probably get another job... But when I posted the test on the GS mastering forum, nobody seemed interested in either taking the test or posting their results.

In any case, we'd need more honest trials from more people.

Martin


16 bit vs 24 bit

Reply #121
Could this whole situation be simply from Martin's poor-quality audio card causing some kind of non-linear distortion that makes these results actually audible when normally they are not?
Could be. I tried to remove the most obvious possibility by taking the 16-bit version, converting it back to 24-bits, and filling up the bottom 8 zeros with noise. Martin could still ABX this against the 24-bit original (it's a couple of pages back - not sure the ABX stats are sound though).

Any other ideas how we could remove this possibility?

Cheers,
David.


16 bit vs 24 bit

Reply #122
Based on that statement, I've come to the conclusion that you're pushing some weird subjectivist agenda. As no one with decent equipment can validate your results (and those who try and fail are not liable to post and say they did) and your results seem to be primarily on a consumer-grade sound device, the conclusion I come to is not that you're hearing a difference between the samples, but rather hearing a difference in your hardware's rendering of the samples.

I'm not gonna argue this any more, I'm just out.

16 bit vs 24 bit

Reply #123
Here's an interesting question (for the 16-bit version at least): Does any similar (but completely uncorrelated) noise make it sound worse, or is it only dithering+truncation that makes it sound worse? In other words, is noise_A a special "bad-case", or not?


Not sure I follow you 100% here, please help me understand, but you make the assumption that noise_A is only... noise?
I mean testing what you have been saying...

Dither+truncation adds noise of a certain character (e.g. white, -90dB RMS), but it is correlated to the original signal in some way (not a way that most people believe is audible, but that's the point!).

You could have noise with exactly the same character (white, -90dB RMS), which has nothing to do with the original signal.

It would be interesting to know if both can be ABXed against the 24-bit original, or if the latter cannot. That would determine whether what you can detect here is any old noise, or a specific type of noise.

Cheers,
David.

16 bit vs 24 bit

Reply #124
As far as audible on a good system, I would go as far as claiming that any full time mastering engineer not able to hear the difference between the 24 bit source and the 16 bit on his own setup should probably get another job...
Sighted tests don't count around here. There are thousands of examples of "obvious" differences that vanish in a double blind test. Thousands of examples of things that "everyone knows" which, when it comes to it, no one can actually prove.

Sadly, the audio world is full of them.

Also sadly, many people who claim to hear such things refuse point-blank to do ABX tests.

As far as I know, you're the first person to ABX 24>16-bit conversion at a low level. That's kind of extraordinary.

Cheers,
David.