16bit vs 24bit, Rubbish or Truth? Options

[Bourne]

-

This post has been edited by Bourne: Apr 2 2008, 19:26

[AndyH-ha]

Aside from the fact that it is almost always easy to tell live from a recording, no matter how good the recording, comparison of live vs recording of any particular bit depth is not only unnecessary, it is completely irrelevant to the question of whether there is any audible difference between a 24 bit recording and the same properly converted to 16 bits.

The question is simply can anyone actually tell the difference, merely by listening -- when they have no information except what they hear to tell them which is which. We are not talking about difference in equipment or listening environment, these have to be identical while listening to both versions. We can certainly consider the question of how good the equipment and environment need to be to allow a difference to be perceived, if such perception is possible, but that is not the issue itself.

The claim that only raw 24 bit recordings are clearly different is an interesting one. Just why a difference, audible when recorded at 24 bits, would disappear once the tracks are mixed, etc. (still at 24 bit) is unclear, unless you are talking about the mistreatment most pop and rock music is subjected to on it way from recording to distribution. Quite a few people have pointed out that SACD and DVD-A releases sometimes receive mastering that makes them better, and clearly distinguishable from, the CD version of the same recording. However, as the tests in the paper demonstrated, those differences are not lost or altered when that SACD/DVD-A recording is converted to CD specs.

For five years or so, whenever this subject comes up, in this and a number of other audio oriented forums, I have issued the challenge for anyone championing >16 bits and/or >44.1kHz, to provide a sample: something I can convert or resample (properly) and then identify in a correctly done blind ABX test. Only a real music recording is applicable; it isn’t as hard to make up a sample with test tones.

I don’t say that such recordings do not exist, only that I, and at least most of the other people on the planet, have yet to hear one. So far, every time I’ve made this request, the people arguing in favor of greater bit depth and/or sample rate have just gone away. (There is one thread, in this forum, where a poster claimed to be able to ABX just about any change whatsoever to some 24/96 samples, but I don’t believe anyone else heard what he claimed to hear. (There were also some questions raised about his equipment) The thread got so convoluted that I stopped following it, but I don’t think he convinced anyone.) (Also, that was just about sample rates, not bit depths).

Now I acknowledge that my inability to hear a difference, should a test file become available, does not mean no audible difference exists. The sample has to be available for wide ABX testing, to find out if everyone is as deaf as me. Many people who visit this forum are experienced in doing ABX testing.

Three to ten seconds should be adequate, I see no need for an entire track, but will consider any reasoned argument as to why a greater duration might be necessary to reveal a difference. My preference is for 24/44.1 since even a few seconds of audio is tedious to download over my dial-up line, but a restriction to CD spec sampling rate isn’t a requirement. I, and probably most people, can handle 48kHz, 88.2kHz and 96kHz.

This post has been edited by AndyH-ha: Jul 7 2008, 06:00

[ccryder]

Aside from the fact that it is almost always easy to tell live from a recording, no matter how good the recording, comparison of live vs recording of any particular bit depth is not only unnecessary, it is completely irrelevant to the question of whether there is any audible difference between a 24 bit recording and the same properly converted to 16 bits.

Two points to make on this.
#1, if you don't know what you're supposed to be able to hear in that recording, i.e. if you don't have a live source (or preferably a live analog source without the A/D stage) to compare to the two recordings (consider what I'm talking about to be not an ABX comparison, but rather an ABX-C comparison, where C is the live analog source), then it changes your perception of what kinds of differences to expect from the outset. It's been my experience that different kinds of music, with different harmonic and reverberant content will have an effect on how drastic the differences between 16 and 24 bit will end up being. I start with the premise that the most valid test should be to determine which one produces a better sounding facsimile of the original. I maintain that without a reference to a live source... without a glimpse of what should be possible to hear in that passage, your brain may have nothing specific to look for when you're then tasked to look for differences between A&B. However, (for instance) if you know how much reverberant content the original source contained, I propose that would educate your ear for that particular test, and would change outcome of such a test.

#2, take 2 recordings, both recorded conservatively so as not to risk ever clipping or hitting 0dBFS for more than 1 sample. Say, -24dBFS is the highest peak on a simultaneouly recorded 24bit and 16bit recording. I don't care how much dither and noise shaping you add, you're not going to end up with more than 12 significant bits on a 16 bit recording, but with good quality analog components & A/D, you will still end up with 16 significant bits of audio on the 24bit version. The difference in listening back is effectively the difference between a 12bit recording and a 16 bit recording (noise floors notwithstanding). Can you hear that difference? Do you understand why you might hear that difference? Now, take an instance where there's one transient spike +12dB higher than the rest of the recorded music, and you happen to know exactly when (in advance) that spike will occur. So knowing that, and not wanting to ride the record levels to as to manually compress dynamic range, you leave the levels set in such a way that the majority of the peaks are at -12dBFS, and you allow room for that one spike to take it a fraction under 0dBFS. The majority of that recording in the 16 bit realm is still getting a "14bit treatment", and even then only at its loudest point that it gets that amount of resoution . And what about that 24-bit version? Still gets a full 16 bits or more of significant dynamic range. Think you can hear that difference? I'll make the bet again and again that any decent audio engineer can tell the difference between a 16bit recording and a 12-14 bit recording. And therein lies the reason a live, unmastered, unadulterated 24bit recording will, more often than not, be obviously better sounding than its 16bit counterpart.

I should say after reading recent posts (and even before reading them), that I agree with about 95% of the sentiment that higher sampling rates are largely indiscernible. I think it is possible that there are special situations when it might be possible to reliably discern the difference between 24/44.1 and 24/96.... but I would probably limit such situations to where a single acoustic instrument containing high frequency harmonic was being captured with *near coincident* stereo mic patterns like ORTF. It would be that type of situation that I would seriously doubt ever entered into the test equation performed by Moran and Co. Asked if I bother with 96kHz, I would tell you that only if I cared very deeply about the recording for my own personal music, or was being paid specifically for a 24/96kHZ, would I use 96kHz to record.
The rest gets the 48kHz treatment. Generally speaking, for mastered, fully produced music, I agree that 96kHz for playback is pretty much a waste of space and CPU power. Not having read Moran & Co.'s paper, but from the articles mentioned above, it would appear to me that the general focus of the test was to prove that higher sampling rates are a waste, and not greater wordlengths. Like I said, most of the time, and for the overwhelming majority of both listeners and listening material, I agree with their findings.

However, wordlength is an entirely different matter.
Bottom line: I believe unmastered 24-bit audio compared with unmastered 16-bit audio from the same source has the potential to indicate vast audible differences in the timbre and decay of both harmonic and fundamental content (as well as reverberant content) over time. Most people don't prefer to listen to unmastered, yet well captured recordings. However, there are some that do, like me, and given what I've mentioned above, for those people, it's night and day. However, crush the crap out of the 24bit recording, and master it to 16 bits with noise shaping and dither, and then compare the two, and you probably won't hear much of a difference.

Back when I did a lot of concert recordings (in those cases, unpaid jobs), there were many times I, being the only one using 24-bit technology at those venues, created a decent sounding recording, while everyone with 16bit technology ended up with noisy rubbish, solely because the majority of the content was far below -0dBFS.

There are a whole host of live, unmastered 24bit recordings on archive.org (of varying qualities based upon mic placement from the source, mic patterns, analog stages, and A/D stages).... some of which I personally recorded, and have been available publicly for the past 6 years or more. Anyone who wanted to do a test for themselves to hear the difference (provided they had playback capability with enough dynamic range) can pull down one of those recordings and do their own experiments. I suspect that most who say they don't hear a difference never tried to listen to this kind of source material.

So for anyone who prefers this kind of listening to the nightmare creations of audio engineers designed specifically to sell more records, there is no argument.
Anyone's perception of the value of 24bit recording *and listening* is simply a matter of perspective and expectations.

I recognize that the majority here doesn't listen that way, or to that kind of music, but I just thought perhaps a glimpse of thinking outside the box might put the argument of 16bit vs 24bit into better perspective, for both recording, and listening. Minimally, any argument in that vain needs to be first qualified by the type of source material and listening attitude of the participants.

-DH

[euphonic]

#2, take 2 recordings, both recorded conservatively so as not to risk ever clipping or hitting 0dBFS for more than 1 sample. Say, -24dBFS is the highest peak on a simultaneouly recorded 24bit and 16bit recording. I don't care how much dither and noise shaping you add, you're not going to end up with more than 12 significant bits on a 16 bit recording, but with good quality analog components & A/D, you will still end up with 16 significant bits of audio on the 24bit version. The difference in listening back is effectively the difference between a 12bit recording and a 16 bit recording (noise floors notwithstanding).

This 12- vs 16-bit comparison is valid only if the listener deliberately jacks the volume way up during quiet bits and turns it down during peaks, i.e. an artificial scenario. Assuming the volume is kept constant throughout the recording, the total resolution is still fully 16-bit. I don't think 16/44.1 can be deemed insufficient based on such manipulation of playback done for the sake of hunting for artifacts.

This post has been edited by euphonic: Jul 7 2008, 23:57

[ccryder]

OK, Gonna try once more to get some of you to understand what I'm talking about.
If you don't get it after this, I would venture to say you may never, or you don't want to get it, or you just don't like my explanation.

The overriding concept seems quite simple, and would appear to me to render the whole argument about audible differences moot.

AndyHA implied that any perceivable differences between 24bit and 16bit audio should be resolved with dither and noise shaping, and that if I'm hearing differences, it's because I didn't add "properly applied noise shaping." (I love that phrase "properly applied..." it says so much, as if there's an improper way to apply it, other than to apply it twice, or to perform additional post-processing afterward. Anyway....)

Let's start with an undithered 16bit quantization from a high-grade 24bit source. (Sample rate is irrelevant for the topic of discussion, as long as it's 44.1kHz or better.) Most wouldn't argue against the idea that this quantization noise is audible, and would therefore suggest that dither be added to reduce the perception of quantization noise. However, dither itself
adds some amount of broad spectrum noise to the signal, and raises the noise floor. Not much argument there. So now we have 2 kinds of noise that can possibly appear on a 16 bit recording, and I don't hear anyone debating whether or not either of these noises are detectable. It would appear universally accepted that indeed these noises can be heard.... Maybe not by everyone, maybe only by people with the right playback equipment and a trained ear.... whatever. Nobody argues that, and in fact, many here seem to feel comfortable implying ad infinitum that the panacea end-all-be-all solution to this is noise shaping.
If only it were that simple.

So now we've established that there is a necessity for noise shaping on a dithered signal, because without it, noise is more easily detectable. Great. Yes, I can hear that noise, and so can many others. Bring on the noise shaping! Oh but wait.... ummmm.... which noise shaping algorithm should I use? Hmmmm. Geez, there are so many. Surely they're all not the same, or there wouldn't be so many. What are the differences between them all?

Some noise shaping is available in hardware, and some in software. All of the manufacturers of noise shaping products/algorithms make claims of the perception of increased dynamic range to varying degrees. Many of the manufacturers make available more than one algorithm, with different perceived dynamic range claims... Well, why wouldn't I want to just use "the best one?" We can make the incorrect layman's assumption that the best one is the one that gives us the perception of the greatest perceived dynamic range, based purely on the manufacturers claims. Or you can ask an audio engineer how they choose the "best one."

(side note: Amazing how many will bash audio companies and producers of high end audio equipment, labeling them mere opportunist marketeers who use vague language to push unnecessary products on unsuspecting customers... But somehow these same people are more open to blindly accepting the noise shaping manufacturer's claims of increased perceived dynamic range numbers as gospel.... perhaps next someone will try to invalidate that statement by trying to convince me now that every noise shaping algorithm ever invented is just some pure audiophile marketing bull, and that it's not necessary... but I digress.)

Anyway... So we're starting from a well established point where it has been determined and widely accepted that quantization noise without dither can in fact be heard. Then you take the next step and consider that it is widely accepted that the best solution to resolving the quantization noise issue is to apply dither. OK. Fine. Now, I don't hear anyone making the argument that this dither noise cannot be heard. If anyone were to make that argument, they would be arguing against every noise shaping algorithm producer out there, and it is generally accepted that noise shaping has much value in this context. And of course, we can see that AndyHA fully believes in the power of noise shaping to reduce audibility of the added dither noise. And so do I. Note that I said *reduce* the audibility, and not eliminate.

Different noise shaping algorithms attempt to shift noise to either specific fixed frequency bands where the ear is less sensitive according to a fixed curve, or to dynamically modify where the noise is shifted based upon the signal content at any point in time in an attempt to artfully allow the noise to be masked. I believe Apogee makes this "dynamic noise shaping curve" claim with their UV22 product. In any case, the choice of a noise shaping algorithm is a personal one. Why? Because the different noise shaping algorithms have different curves, applying different amounts of "energies" to different frequency bands, and ultimately affecting different people with different hearing sensitivity curves. Simply put, just about every noise shaping algorithm is considered to impart a certain coloration to the signal. (Of course there are manufacturers that think they've come up with the best, non-coloration algorithm, and can't wait to tell you about it). Some might shift the noise to the 10kHz area, some to the 12kHz area... some put a "clump" of noise around 4kHz and another clump at 18kHz. Some attempt to push all of the noise toward the Nyquist limit, but do so with a smooth slope starting at 18kHz and increases the level of noise on higher and higher frequency bands. None of these numbers I mentioned are specific to any one algorithm--those numbers are just examples of how one noise shaping algorithm might differ from another. So as it turns out, the choice of noise shaping is somewhat of a compromise. If we have more than one algorithm available to us, we choose whichever algorithm we feel gives us the most pleasing results on a specific recording and specific type of music. Some people may find they like the sound of a noise shaping algorithm that pushes all of the noise to higher frequencies on a piece of music that already has a lot of very significant high frequency content, because they like the particular coloration the noise adds. Others might prefer that the noise was located somewhere else in the spectrum where it would be better masked by more dominant frequency content. The choice of algorithm is a personal one, and there can be many combinations of algorithms and signal type. Some manufacturers will indicate that X algorithm is probably best for quiet classical music, and Y is best for rock and roll. Some will say that Z is best if you ever plan to run any more post processing on the signal (something I'd never do anyway).

Needless to say, all of this implies that the results of the application of different noise shaping algorithms will impart varying degrees of coloration (or perhaps even to the degree that an individual, i.e. *not everyone* might not perceive the coloration at all). It is an artful dance that the engineer must make to choose which one he thinks makes the music sound the best. You have hearing curves to consider, the music has a particular EQ, and you try to match the noise shaping algorithm to strike a balance (compromise) between unpleasant coloration and unpleasant noise for the average Joe.

Great. So we know we can hear differences between dithered and non-dithered audio. We know that we can hear differences between noise shaped audio and non-noise shaped audio. And we know that a trained ear can hear different colorations of different noise shaping algorithms.

So given all of that, why is it that people think you can't hear the difference between a non-colored raw 24bit recording, and a dithered and/or noise shaped (colored) 16bit rendering from that same 24bit source?

All you folks championing the seemingly limitless merits of noise shaping in a very generic sense as the "difference eliminator" are way off-base. Nobody mentions the type of music, the type of dither, the type of noise shaping, and the type of listener. They just love to say "I don't believe anyone can hear a difference." It is statements like that that are just plain hogwash. This is pure logical deduction. If you (or *anyone else if not you*) claim to be able to hear the difference between one noise shaping(NS for short) algorithm and another, then it follows that you can hear the differences between NS audio and non NS audio, and dithered only, and non-dithered audio, that pretty much covers the full spectrum of discernibility. That's it. Period. No need to belabor the whole 16-bit vs 24 bit difference argument. The differences cannot be boiled down to any one thing for everyone, but for many, there are audible differences. Accept it, and get over it. Whether or not I (or anyone else) am more perceptive or not to these types noise and colorations is not for anyone to argue. If you want to talk about specific combinations of recordings, levels, listeners, technology, and experience, then perhaps there is a basis for discussion and specific analysis of perceptibility for that combination only.

Alas, as I did many years ago, I performed another test myself tonight. Yes, it was a blind test... or blind enough anyway (can't wait for the reaction I get for this one, but don't expect a response from me). Very simple. I took a 24-bit48kHz recording I recently made of a friend's classical guitar recital. It is a very dynamic recording, with signals ranging from "totally in your face" and yet not clipping, to extremely quiet--sneezing, breathing, sniffling, coughing squeeky-chaired audience notwithstanding. The kind of recording where you can hear people whispering from the back of the room intelligibly if you turn up the amp.

It was recorded with a stereo pair of Schoeps CMC64V, ORTF, onstage, about 4 feet from the performer, 2' off the floor. A Grace Lunatec V2 preamp (130dB dynamic range, per specs), Benchmark AD2402-96 A/D converter (118dB dynamic range, per specs), and some expensive (I'll make no claims of them being particularly mindblowing at this time) mic and signal cables were employed. I played it back in Winamp with no altering plugins or processing of any kind, which fed the signal to a Soundscape Mixtreme PCI 24/96 PCI card on a known bit-transparent audio workstation. The sample rate set at 48kHz internal to match the 24/48 recording. The Soundscape card is particularly cool, because it allows me to insert live VST plugins into the signal chain. So I chose to insert the old standby plugin... Waves L2 Ultramaximizer. I disabled all processing in the plugin, other than the dither section. No gain changes, no limiting, no ARC... just 3 settings: bit-depth, dither algorithm (type 1 and type 2), and noise shaping (none, Moderate, Normal, and Ultra. There is an A/B button in the plugin that allows one to set up 2 different configs. I set "A" to 24bit, no dither, no noise shaping. I first set "B" to 16bit, Type 1, and Moderate. I started the music playing, put on my Sennheiser HD650 headphones, plugged them into my Benchmark DAC1 D/A, positioned my mouse over the A/B button, closed my eyes, and clicked the mouse over the button some random amount of times in rapid succession, with specific intention to lose track of which one I was on... A or B.

My goal? 3 goals, which many it seems would like to confuse. #1, can I hear a difference? #2, can I pick out the 24-bit version consistently? and #3, can I pick the best sounding one consistently. Truth be told, I didn't start out with all three goals in mind. #2 implies #1, so initially, #2 was the goal. #3 was something I fell into and had to train my ear more to discern.... but that test was not quite solid for reasons that can be inferred from the rest of the test results below.

Anyway, So there I was clicking the A/B button, over and over again, not knowing which was which, with my eyes closed, trying to pick out the 24 bit recording from the 16-bit, type 1 Moderate noise shaping. If I heard something in the music that gave me the impression that whichever setting I was on was definitely the 24-bit setting, I opened my eyes, wrote down which setting I was on, and started the process over again clicking the button with eyes closed until I could effectively randomize which setting I was starting on. I flipped back any forth many times while I listened. Sometimes it took 20 A/B flips or more in rapid succession and the right music passage to combine in time that gave me the immediate "aha!" moment where I thought I was sure which one was the 24 bit material. I did this whole round 20 times for each possible noise shaping algorithm, which I thought would be a decent sampling of attempts to start with. On that first test, I properly selected the 24-bit recording 18 out of 20 times. The next test, normal noise shaping, type 1 dither, and the results were 19 out of 20 times correct. My ear was also becoming trained in hearing the differences. 3rd test, type 1, ultra noise shaping, and the results were 14 out of 20 correct. The ultra test was definitely harder, and I had to ask myself why, and I discovered that I just had a personal preference for the way that particular coloration enhanced my perception of what was better sounding for this recording, and my brain more often made the erroneous assumption that the better sounding one had to be 24 bit. Being familiar with the curve the ultra algorithm used, I basically trained myself to recognize the increase in high frequency energy which is characteristic of that algorithm, and went back and did the 3rd test again, and the results improved to match the other tests at 18 out of 20 correct. While I did the same tests with type 2 dither, the time that it took to discern the differences continued to go down, and my accuracy remained the same or better. I also did the test with no noise shaping, and it was a no-brainer. I could too easily tell that there was dither noise every time the levels died down, and found it frustrating that it was this noise that continued to tip me off before other factors.

Is my test the exact same as a double blind ABX test? Not quite. I didn't have the ability to make it so that the A & B's were randomly not actually switched in my test. Considering that I switched back and forth with closed eyes until I knew for sure which was which, I feel that the methodology well served the goal of answering the question "can I hear the difference?" I also didn't have the ability to mask the results from myself, since I performed the test alone and had to write down the results as I went along. I don't believe that to be a significant problem, since the next sample was otherwise an independent trial, and there was something to be learned from that as well.

What did *I* learn from this? I learned that my assertion that the ear can be trained to hear differences is valid, and that there is something to be said for allowing time for the ear to be trained. If the ear could not be trained over any length of time or number of trials to achieve results statistically significantly different, then it might prove all of these assumptions about indiscernability true. However, this was not the case. Further, I learned that what one uses to discern one from another may vary from passage to passage, person to person, and configuration to configuration. Sometimes it was perceived depth of the soundfield due to seemingly more persistent reverberations, sometimes it was specific colorations I detected in certain frequency bands, and sometimes it was the ability to perceive the noise. It was not the same thing each time that tipped me off, but each time I was convinced I was listening to the 24-bit recording for some reason or by some method on some passing passage, I looked up, and saw that I was correct often enough that it wasn't a fluke. The times I wasn't correct, I can only attribute to the specific combination of signal content, level, and coloration that made me think it sounded better in the moment. Finally, I learned that one second it might not be discernible, and in the next second it is.

The final test #3 that was only slightly necessary for me, as it was an afterthought... can I pick out the better sounding one consistently? In all cases except with the ultra shaping, I again chose the 24bit recording with regularity. For whatever reason, the ultra was adding something akin to an EQ that was making it particularly pleasing for that recording, and I continued to be pulled towards that 16-bit configuration for that recording. I know from previous experience that on other recordings I've mastered that the ultra algorithm adds too much coloration in high frequencies to be pleasing enough for me to call it better for those recordings (not apparently so in this one).

Bottom line for me:
24-bit vs 16 bit no dither = easily discernible, most agree
24-bit vs 16 bit with dither, no NS = easily discernible based purely on perceptibility of dither noise alone, most agree
24-bit vs 16-bit with dither and varying algorithms of NS = generally discernible for many reasons that varied from moment to moment, with notable improvement of accuracy over time due to unpreventable ear training. And yet, for some seriously odd stubborn reason, many continue to be in denial of this.

If your can hear differences in NS algorithms for a given audio source, then you can hear the difference between NS'ed and non-NS'ed audio.

Find flaws with my tests? Disagree with the premise entirely? Think what I wrote is too damn redundant (no doubt it is at times)? Then do your own tests, on your own 24bit recordings, with your own high-grade equipment, make an effort to train your ear to hear the different noises, artifacts, colorations (whatever you want to call them) and then let's talk turkey. It's just a silly baseless debate otherwise, typically accompanied with a lack of disclosure of all of the variables that truly matter, a lack of significant hands on experience with 24-bit audio, and completely invalidated by the simple premise that audio engineers every day choose NS algorithms by sampling with their ears which ones "sound the best" for a particular recording.

-DH

This post has been edited by ccryder: Jul 8 2008, 07:58