• mjb over 2 years ago

    Continuing an off-topic discussion from another thread...

    After spending way too much time in the Hydrogenaudio forums, and after making thousands of vinyl rips for nearly 20 years, I'm quite an audiophile skeptic. A conversation about preamp & ADC gear quickly diverged into me challenging claims about whether 24-bit recording offers advantages over 16-bit, at least for vinyl ripping.

    On Discogs and elsewhere, the main claims I see are:

    1. "I can hear a difference and 24-bit sounds better". This claim is often the result of apples-to-oranges comparisons (the clips come from different sources, are at different volume levels, were badly converted, or are getting converted differently upon playback), or it is the result of the placebo effect and humans' strange ability to hear differences where there are none. Repeatedly play the same clip for someone but tell them one is 16-bit and one is 24-bit (or whatever other lie you want to tell) and they may well say that they hear differences, especially if they believe that one should be better than the other. With enough tests, this can be proven to be a random phenomenon, though; one clip will be "better" just as often as the other. To know whether you are hearing a difference requires blind "ABX" testing. Challenges on the Hydrogenaudio forum have yet to produce any ABX test results confirming 24-bit music recordings are distinguishable from 16-bit.

    2. "24-bit allows you to set the recording level lower/less precisely." Yes, but... It's true, the quietest sounds on vinyl (the sound of the empty groove walls rushing by the needle in at the beginning or end of a side) will not be fully captured in a 16-bit recording if you set the recording level too low. Go absurdly low and you lose the quietest parts of music, too. With 24-bit you have more leeway, but you probably have more room than you think in 16-bit (peaks can be -20 dB without consequence, in my experience).

    More importantly, 24-bit does not cure SNR issues. That is, your gear inevitably gives you some electrical background noise (hiss, hum, tones, clicks) which are at a constant, hopefully very low level. This noise level will not be any different in 16-bit than 24-bit. You may actually be picking up even more noise with 24-bit. Reducing the volume level of the recording is just bringing the signal closer in volume to that all that noise, effectively raising the noise level into the audible range. If you're serious about sound quality, you want to maximize SNR by keeping the recording level as high as it can be without clipping or dynamic range compression.

    3. "24-bit is safer for processing/mastering work because it's more accurate. Each transformation of the data introduces tiny rounding errors which can add up to big ones." This is another "Yes, but..." and is where the discussion left off:

    chiz
    [if there are] one or more steps of processing is this not done more accurately at the higher bit depth [...] any noise introduced by the processing is at a lower level

    Sure, but how accurate does it need to be? We can think about possible disasters all we want, but the proof is in the pudding. I've run my own tests with many different settings, and encourage others to do the same. It's a good way to learn about all the things that affect and don't affect the end result, as well as quirks of your gear.

    I've probably read the same things you have, but after doing my own tests, I am way more skeptical that rounding errors during processing are ever something to actually worry about. Compound/additive error from DSPs is easy to imagine but difficult to actually produce in DAW software because the software uses 32-bit (or even 64-bit) floating-point samples internally, even when editing integer files.

    Do what I did: try applying a hundred or more transformations to a 24-bit source and a truncated 16-bit version of same; if your test turns out like mine, the 16-bit result will be sample-for-sample identical to the 24-bit result truncated to 16. Maybe if you were to save, close, and reload the file after each transformation, you'd randomize a lowest bit or two, but I would be amazed if you could ever produce audible differences that way, at least with the number and type of transformations people do on vinyl rips.

    ...

    So I am quite satisfied that 24-bit is never needed for recordings from any analog medium, except for peace of mind.

    Nevertheless, a friend once told me that in his experience, 24-bit vinyl rips can sound better not because they are 24-bit, but because the people who make them tend to use higher quality gear and be more conscientious overall. I conceded on that point!

    I love the taste of crow as well as humble pie, so please, run your own apples-to-apples tests, comparing only changes in bit depth (not sample rate), share your methodology, and show us the ABX logs as proof that you can hear the difference. You may instead want to start by exploring http://www.audiocheck.net/ instead... it's quite no-nonsense through-and-through.
  • chiz over 2 years ago

    mjb, thank you for taking the time to answer my question so comprehensively.
    I must admit that I slipped into the habit of capturing and processing at 24 bit due to the theoretical advantages without any practical testing as to whether these could be perceived.
    Thanks for the education!
  • loukash over 2 years ago

    loukash edited over 2 years ago
    mjb
    You may instead want to start by exploring http://www.audiocheck.net/ instead

    http://www.audiocheck.net/blindtests_16vs8bit_NeilYoung.php is fun!
    Current score: 8/12 (67%) — Confidence : 80.61%

    There you go… :)

    Tested in AKG K271 MkII headphones.
    The difference is there, but it's very subtle at this loudness level for my 50 years old ears.
    However, since the last 3 tries were all wrong, I guess I got primarily bored by the song itself. :D
  • mjb over 2 years ago

    mjb edited over 2 years ago
    I tire of Neil Young pretty quickly, too.

    8-bit has quantization noise only 48 dB down. If the music were a solo acoustic guitar and voice, or the fade-out, it'd have some quiet "space" in it and you would probably hear the noise. The challenge with this clip is that it is pretty dense and full of noisy instruments already, so there's a lot of masking happening; it is not easy to discern the relatively quiet added noise from the loud music and the noise that it's in it already. I cannot hear any difference, myself.

    A clever thing about ABX testing (which is what those blind tests on audiocheck.net are) is you're forced to choose one or the other, so it's like evaluating a coin toss. If you randomly guess or are otherwise unable to reliably hear a difference (regardless of why), your success rate will approach 50% over time. In other words, people who can't hear a difference will still get 6 out of 12 right. A lot of people will get 7 out of 12. You have to outperform them by a lot if you want bragging rights!

    https://en.wikipedia.org/wiki/ABX_test#Confidence explains that 95% confidence (for n tries, n/2+sqrt(n) correct) is what's needed to be considered "stastically significant". For 12 trials this means you have to get 10 right. This is what audiocheck.net's blind tests look for (if you get 10/10 or 10/11, the test ends right then).

    Would getting 10/12 be proof that you can hear a difference? Well, just by chance, about 1 in 8 random guessers will get 8 out of 12 right. 1 in 19 will get 9 out of 12; 1 in 62 will get 10 out of 12; 1 in 341 will get 11 out of 12; and 1 in 4096 will get 12 out of 12. So no, if you got 10 right it's not definitive proof, but you will have beaten the odds by a wide enough margin for science! Beat the odds several times over several different days and you will truly earn your bragging rights.

    The intro of https://www.fourmilab.ch/rpkp/experiments/statistics.html explain coin-toss probability a bit better than I can, before it gets into calculus-level math which I once understood but which has completely evacuated my brain in the intervening 3 decades. I had to use an online calculator to get the probabilities above.

    The Windows audio player app foobar2000 has a nice ABX tester built-in. You select two files in the playlist, right-click and choose to do an ABX test. It will match their volume levels and walk you through a testing interface. Pretty neat, but the trick is not getting sick of whatever music you choose to compare!
  • loukash over 2 years ago

    mjb
    I cannot hear any difference, myself.

    I thought I'm hearing a slight difference in the cymbal tails: the 8-bit file seems to sound slightly harsher.
    But I can only hear it if immediately switching back and forth. If I have to tell which is which I simply don't "remember" which one was sounding harsher.

    Back on topic:

    mjb
    2. "24-bit allows you to set the recording level lower/less precisely." Yes, but...

    For me, 24-bit vinyl recording is just an intermediate step. I'm not archiving them.
    After all individual editing is done, I simply batch normalize and convert all processed files to 16-bit in one go, in place, using Audiofile-Engineering's Sample Manager (now known as Myriad) workflows. It's just a matter of an extra minute.
  • mjb over 2 years ago

    mjb edited about 1 year ago
    Another simple test and learning opportunity is to just do bit-depth reductions in software and see what you notice.

    Use headphones in a quiet room. Load a vinyl rip into your editor. Set your amp or headphone volume as loud as you can comfortably tolerate hearing the entire song. Experiment with changing the bit depth to 15-bit, 14-bit, 13-bit, and so on. Try different dither options. Compare to the 16-bit original each time. Check all parts of the song, loud and quiet. If there is a fade-out, listen carefully to it.

    My results, listening to a pop song, were that most of the time, 12-bit without dither was transparent (indistinguishable from the original). But in quiet parts and fade-outs, I needed 15-bit undithered, or 12-bit with noise-shaped dither. At lower bit depths, the quantization noise or dither was audible, and I started losing signal in the fade-out.

    At 11-bit, quantization noise was only occasionally audible; if the song I chose had been noisier and less dynamic, I'm confident that 11-bit would be transparent. At 10-bit, however, quantization noise was plainly audible under these optimum listening conditions. Dither helped, and the music was still fully present, but I still would not call the result transparent.

    Quantization noise sounds like white noise at first, but as the bit depth goes down, it becomes increasingly raspy, tonal, and "correlated"—prone to fluctuate along with the music. At the higher bit depths, it was fairly constant, as if someone had overlaid hiss on top, like you'd hear in a distant FM radio broadcast. Still, even going down as low as 8-bit, the music in the normal & loud sections of the song was fully present; it was just in a bed of raspy, fluctuating noise.

    As you may know, dither adds noise, but also changes the correlated noise to uncorrelated. It smooths out the quantization noise so it's constant and much less annoying, making it a kind of hiss that's easier for our ears to "hear through". "Noise shaping" concentrates it in some frequency ranges more than others, allowing the bulk of the noise to be inaudible. Dither also effectively extends the dynamic range, so you can partially preserve the very quietest sounds from the original. It is almost like having 3 extra bits, but at a price of added noise. Dither can also introduce clipping if you're not careful; if you normalized your rip to have 0 dB peaks, you didn't leave any room for adding dither during this bit rate reduction experiment, so watch out for that.

    In Audition, these are the dither settings I used: Triangular (shaped), Neutral (heavy), Crossover 17 kHz (just below the extreme upper frequency limit of my hearing), Strength 50 dB (but reduce this by ~6 dB per bit below 12, and watch for clipping), Adaptive mode off. Essentially this makes the noise level constant and very low until 17 kHz, where it steeply ramps up and then stays flat and heavy at 18 kHz and above. The strength setting is how loud the noise can be when it's above the crossover frequency. The louder you make it, the quieter it will be below the crossover frequency.

    If I reduce the sample rate at the same time as the bit depth, there's less room for the noise shaping to work, and I get more noise across the entire spectrum. However, going from 48000 to 44100, the difference is too slight to change any my observations above.

    By the way, last time I tried to use Audacity (the free-of-charge editor that many people choose to use instead of Audition, Sound Forge, or other non-free, professional DAW software), I was alarmed that by default, it had fairly loud, non-noise-shaped dither enabled for all output. Every time you opened and saved a 16-bit or 24-bit file, it was adding dither even if you didn't make any changes or were just doing simple cut-and-paste edits. You'd think you were just trimming the ends, but when you saved, you added hiss to the whole file. You could change this behavior in the settings, but it was the kind of thing most people probably wouldn't even think to look for. (There's a related bug report going back to 2009.)

Log In You must be logged in to post.