24/192 Music Downloads Are Very Silly Indeed (2012)

I’m not sure how big that market is these days, though. I recently decided to move from a block in the city to a house in the surrounding countryside purely to get a quieter listening environment and let me really enjoy the high-dynamic-range recordings I have – I listen to a lot of classical music, especially the avant-garde with its even greater range, e.g. the Ligeti cello concerto starting from pppppp. Yet even among my friends who are really obsessed with such music, seeking a better listening environment seemed an extreme measure to take.

So, people who a generation ago would have invested in higher-end equipment (not audiophile snake oil, just better speakers) and who would have sought silence are now giving in to listening to music on their phones or cheap computer speakers. It’s a big shift.

https://xiph.org/video/vid2.shtml

So well prepared, so well presented, so little that could be removed without ruining it.

I aspire to do such good demos but always fall so short.

The problem whenever somebody writes about digital audio, is that it is very tempting to hold on to sampling theory (Nyquist limit, etc) and totally discard the problems of implementing an actual Analog-Digital and Digital-Analog chain that works perfectly at 44100Hz sample rate.

I agree with the assesment that 16 bit depth is good enough; even 14 bit is good enough and was used with good results in the past (!). However, the problem is with the sampling rate.

> All signals with content entirely below the Nyquist frequency (half the sampling rate) are captured perfectly and completely by sampling;

Here lies the problem. This is what theory says, however, when using 44KHz sample rate, this means that to capture the audio you need to low-pass at 22KHz. And this is not your gentle (6, 12 or 24db) low-pass filter; no, this needs to be HARD filtering; nothing should pass beyond 22KHz. And this must be on the analog domain, because your signal is analog. To implement such a filter, you need a brickwall analog filter and this is not only expensive, but it also makes mess with the audio, either 'ringing' effects and/or ripple on the frequency response and/or strong phase shifts.

So on Analog-to-digital in 2017, converters should be operating at a higher rate (say, 192KHz), because this makes analog filtering of the signal much easier and without side effects.

Now, for Digital-to-Analog, if your sample rate is 44KHz, you have two alternatives:

a) Analog brickwall filtering, with the problems noted above

or

b) filtering on the digital domain + using oversampling

the article mentions:

>So the math is ideal, but what of real world complications? The most notorious is the band-limiting requirement. Signals with content over the Nyquist frequency must be lowpassed before sampling to avoid aliasing distortion; this analog lowpass is the infamous antialiasing filter. Antialiasing can't be ideal in practice, but modern techniques bring it very close. ...and with that we come to oversampling."

So they are mentioning alternative (b). The problem is that oversampling does not solve all problems. Oversampling implies that the filtering is done on the digital domain and there are several choices of filtering you could use, for example FIR (Finite Impulse Response), IIR (infinite impulse response), etc.

And each one of these choices have side effects...

In short, the problem is that with 44KHz sampling rate, your filter cutoff (22KHz) is too short to your desired bandwidth (20Hz-20KHz). Using a sample rate of 192KHz gives the DAC designer much more leeway for a better conversion. And CONVERSION is the key to good digital sound.

>What actually works to improve the quality of the digital audio to which we're listening?

It is interesting that the author mentions things such as "buying better headphones" (agree), but he never mentions "Getting a better Digital to Analog converter", which is highly important !!

On the other hand, he backs up his claim that "44KHz is enough" with an interesting AES test i was already aware of in the past:

>Empirical evidence from listening tests backs up the assertion that 44.1kHz/16 bit provides highest-possible fidelity playback. There are numerous controlled tests confirming this, but I'll plug a recent paper, Audibility of a CD-Standard A/D/A Loop Inserted into High-Resolution Audio Playback, done by local folks here at the Boston Audio Society.

This is a very interesting paper, and I did have the copy, however the test equipment should be checked. There are systems and better systems. The AES paper cited above had the particularity that the ADC and DAC used were provided by exactly the same machine (a Sony PCM converter), with the same strategy: no oversampling, brickwall analog filters. I can bet (99% sure) that the brickwall filters were identical on the ADC and the DAC on that machine; Murata-brand filters in a package.

The devil, as they say, is in the details.

Consider a dog that lives with a musician who plays, for example, trumpet. The musician plays the trumpet at home to practice, and also records his practices to review.

A trumpet produces significant acoustic energy out to about 100 KHz [1]. When the musician plays live the dog hears a rich musical instrument. When the musician plays back his recordings, half the frequency range that the dog could hear in the live trumpet will be gone. I'd imagine that this makes the recorded trumpet a lot less aesthetically pleasing to the poor dog.

[1] https://www.cco.caltech.edu/~boyk/spectra/spectra.htm

However, all musicians I know use these high-rez formats internally. The reason for that, when you apply audio effects, especially complex VST ones, these discretization artifacts noticeably decrease the result quality.

Maybe, the musicians who distribute their music in 24/192 format expect their music to be mixed and otherwise processed.

I know and understand how incorrect down-sampling from high frequencies can cause distortion in the form of sub-harmonics in the audioable range.

I know about audible dynamic range and how many decibels of extra range 8-bits are going to give you.

I know all this, but I still have to admit: if there's a hi-res recording (24-bit, >48kHz) available for download/purchase, I'll always go for that instead of the "regular" 16-bit 44.1/48kHz download. I guess escaping your former audiophile self is very, very hard.

Anyone else want to admit their guilty, stupid pleasures? :)

If we keep that in mind, then the setup of the problem changes significantly and most arguments made here do not apply to the download itself.

I don't hear a difference between various rips that are 16/128 or 24/192, but I have noticed a difference listening to the Blu-ray version of it (which is many-many gigabytes in size). It is a definitely interesting experience, but the way I can describe it is as an absolute absence of noise.

Every single version but this one exhibits noise at the start (the heartbeat sound) as the sound goes from very quiet to loud.

But, to be fair, it could just be different masters.

That explains why some people can tell the difference between 44.1 khz and the higher-resolution sampling rates. It also means that the ideal audiophile sampling rate is somewhere between 58khz and 60khz, not 192khz.

Say I've got an 18kHz wave and a 9kHz wave and the 9kHz wave is ever so slightly out of phase. Then imagine there are 10 different waves under 20kHz all interfering with each other in different wages.

Is it still possible to reproduce everything accurately?

And on bit-depth and dynamic range: Given that much audio doesn't use the full range available to it, wouldn't higher bit depth increase the fidelity in the range the audio does fill? The article talks about the bit depth and range only in terms of the maximum volume but what about fidelity? What's the minimum difference in volume the human ear can hear?

The higher frequencies (hi hats for example) are mushy and sound like they are warbling, and the sound just generally has a lack of depth. What is the explanation for this, or am I just imagining it?

It's via bluetooth, and the source is Pandora. I even googled Pandora and pitch change, and found some forums with folks discussing bluetooth causing this to happen, which has never been my experience.

It occurred to me that slowing the playback (eg, causing the pitch to sag) could be a fantastic way of dealing with a connection that's too slow. It would be far better than stuttering, even though many of us would find the pitch change really annoying. The instructor and my wife simply can't hear it, or say "I thought that was just part of the song".

Anyway, has anyone ever heard of this happening? Do certain products do it? Were the Pandora forums wrong and this is actually a Pandora problem? Of all of the bluetooth problems I've had in my life, I've never experienced this before, so I tend to lean towards it being a data rate issue on the cell connection and Pandora slowing the music down.

I would expect the samples to be interpolated before the final DA conversion takes place, so no extra ultrasonics should be involved. And there would be filters for them anyway, we still have to use them for 44.1, 48, 96 etc.

Do you not hear anything? Yeah... all that sound is what amounts to lost fidelity when you down-sample. You can't just argue on bit-rates alone... It's about having head room in the mix and room for more fidelity. Sure, if you mix shit badly, you can't hear the difference, but that's missing the point entirely.

It's not nearly as bad as the time someone tried to tell me that Opus was an adequate audio file format for music, but still... frustrating.

I know I don't have super-human hearing, and I can easily hear the difference. Just because everyone can't hear the difference, or can't tell that they can hear the difference, doesn't negate the fidelity loss of lower bitrates.

Further, given a large enough speaker stack, it's not about what you can hear any more, it's about what you can feel.

And never-mind the benefits of a 24bit DAW pipeline... Hello, low latency?

I only use higher bitrates when creating music so when I mix or record I don't need to make the levels go almost into red to get a hot enough signal, I can do all kinds of changes like slowing down or up, distortion, EQ, compression without losing any perceived detail in the end. If I record on 16 / 44.1 then start manipulating the sound I start losing detail immediately.

But in the end more than 20 / 88.1 won't do much for you. 16 / 44.1 might be a bit less than ideal for music with a lot of dynamics but it's absolutely fine for most purposes.

Unfortunately I lost a lot of hearing through concerts, headphones, and traffic, but I can still tell the difference between an MP3 and CD-quality music much of the time.

And the "44.1KHz is enough" or "48KHz" is enough" people are, sadly, kind of dumb.

How do I know? Because I was dumb, too.

Being a coding/math/audio/video badass, after a few years of industry experience, I rattled off some mouthy kid quip, saying "well, I don't know why we do 24/96, since nobody can hear above 20K anyway..."

And a very talented, very knowledgable, and generally reserved engineer suddenly perked up one eyebrow and said, incredulously "because temporal and frequency response are inherently linked..."

That look was in 2001, and I still remember that feeling of dread sinking in, realizing that I had no idea what he was talking about, no concept of why that would matter. I knew about Nyquist and could write a quick FFT, but he'd spent four years getting a degree in pure audio engineering at the most selective program in the country. That look, which I completely remember today, was like a deeply disappointed parent after a kid has just been bailed out of jail, from one of the nicest engineers I know.

It was late, I was brash, and I pressed on, asking him what he meant. "What's a transient look like, spectrally?" he asked. He waited for my blinks to make audible sounds (due to the apparent hollowness of my head), then he asked "and how many channels of audio do we listen to?"

He watched me stand there, like a doofus, for what seemed to me like several minutes (probably 5-10 seconds), then he went back to coding.

It didn't hit me until weeks later, and I didn't really internalize until years later, that he was hinting at inter-ear phasing the other facilities of our auditory systems besides frequency response. Years later, I read up on Georg von Békésy's incredible work (including positional acuity), and I worked as a tech lead at Digidesign on the first-generation Venue live sound system which operated at 48KHz but with incredibly low latency (processing steps were 1, 2, or 16 samples) due to the requirements of, for instance, vocalists using in-ear monitors.

Along the way, I ran across Microsoft engineers who thought that ~10-20ms inter-channel timing consistency would be okay in Windows Vista (it wasn't), conducted blind tests between 96KHz and 44.1KHz audio (for people who were shocked to immediately notice differences), came across plenty of hot-shot kids who said exactly the same kind of stuff I'd said, and saw postings from xiph making a mix of valid and grossly sophistic arguments ranging from "here's how waveform reconstruction from regularized samples work" (good) to "audio equipment can't even capture signals beyond this range" (dumb). At times, I thought about setting up refutation articles, then I realized, like many, that I had actual work to do.

Von Békésy's work points to positional inter-ear phase fidelity of roughly 10µs. What's the sampling interval at 44.1KHz? >22µs? Good luck rebuilding that stereo field at 44.1...

The trick is that there is a really serious diminishing return on audio sampling rate. 4KHz to 8KHz is enormous... 8KHz to 16KHz is transformative... 16KHz to 32KHz is eye-opening... 32KHz to 48KHz is appreciable... 48KHz to 96KHz is.... pretty minor, especially in the age of crappy $30 earbuds, streaming services delivering heavily compressed audio that will be crammed into a BT headset that may or may not be connected with additional compression, and all of the convenience that those changes bring. You may detect it in some audio if you're really listening, if you know what to listen for, and it may present advantages in system design (converters, processing, etc). From a data-rate perspective, the low-hanging fruit has already been picked.

But people who smugly say that there is "no difference", that audiophiles are buying "snake oil", are letting their ignorance show, and that's including that kid that I was, 16 years ago.

I've since moved out of pure pro media to consumer devices, where precision takes a back seat to the big picture a lot of the time. When discussing an audio fidelity multi-channel problem with a possible vendor last year, I expressed my concern about the inter-channel timing assurance slipping from 1µs to 50µs in product generations. "Depending on the sampling rate, that's several samples of audio", I said.

A very senior engineer on our side (Director equivalent) quickly admonished me, saying "it's microseconds, not milliseconds", to which I said "I know... Which is why it's several samples, not several thousand..."

From the look on his face, I'm 100% sure that he didn't understand me at the time, but I hope he put it together eventually.

In the end, the industry has moved in the opposite direction of 24/192 for a long time. If we can get back to normalization of CD quality audio, I'll be happy.

Monty is wrong. To cover the range of human listeners, the required specs even through use of very insensitive double blind testing (which is geared to substantially indicate the PRESENCE of a difference between examples if that's present, and does NOT similarly indicate/prove the absence of a difference with a comparable degree of confidence: that is a horrible logical fallacy with realworld consequences) are more like 19-21 bit resolution at 55-75K sampling.

Beyond this, there's pretty much no problem (unless you are doing further processing: I've established that quantization exists even in floating point, which a surprising number of audio DSP people seem not to understand. There's a tradeoff between the resolution used in typical audio sample values, and the ability of the exponent to cover values way outside what's required)

That said, it is absurd and annoying to strive so tirelessly to limit the format of audio data to EXACTLY the limits of human hearing and not a inch beyond. What the hell? I would happily double it just for comfort and assurance that nobody would ever possibly have an issue, no matter who they were. Suddenly audio data is so expensive that we can't allow formats to use bytes freely? That's the absurdity I speak of.

Our computers process things in 32-bit chunks (or indeed 64!). If you take great pains to snip away spare bits to where your audio data words are exactly 19 bits or something, the data will only be padded so it can be processed using general purpose computing. It is ludicrous to struggle heroically to limit audio workers and listeners to some word length below 32 bit for their own good, or to save space in a world where video is becoming capable of 1080p uncompressed raw capture. Moore's law left audio behind years ago, never to be troubled by audio's bandwidth requirements again.

Sample rate's another issue as only very nearby or artificial sounds (or some percussion instruments, notably cymbals) contain large amounts of supersonic energy in the first place. However, sharp cutoffs are for synthesizers, not audio. Brickwall filters are garbage, technically awful, and expanding sample rate allows for completely different filter designs. Neil Young's ill-fated Pono took this route. I've got one and it sounds fantastic (and is also a fine tool for getting digital audio into the analog domain in the studio: drive anything with a Pono and it's pretty much like using a live feed). I've driven powerful amplifiers running horn-loaded speakers, capable of astonishing dynamic range. Total lack of grain or any digital sonic signature, at any playback level.

My choice for sample rate at the extreme would be 96K, not 192K. Why? Because it's substantially beyond my own needs and it's established. I'm not dissing 192K, but I wouldn't go to war for it: as an output format, I would rather leave the super high sample rate stuff to DSD (which is qualitatively different from PCM audio in that the error in DSD is frequency-sensitive: more noise in the highs, progressively less as frequency drops).

Even with DSD, which is known to produce excessive supersonic noise even while sounding great, the scaremongering about IM distortion is foolish and wrong. If you have a playback system which is suffering from supersonic noise modulating the audio and harming it, I have three words you should be studying before trying to legislate against other people's use of high sample rates.

"Capacitor", and "Ferrite Choke".

Or, you could simply use an interconnect cable which has enough internal capacitance to tame your signal. If you have a playback system that's capable of being ruined just by 192K digital audio, your playback system is broken and it's wrong to blame that on the format. That would be very silly indeed.

I hope this has been expressed civilly: I am very angry with this attitude as expressed by Monty.

Edit: And the linked "Show & Tell" video is a great way to get some "intuition" about the sampling theorem. https://video.xiph.org/vid2.shtml

If you look at the angular resolution of the eye, unless you are sitting very close to the screen, you can't resolve 4k video.

So, you have to up-sample the songs to high rates with less bits, like 1bit to 6bits, then do the conversion, and get the best SNR you can.

In this sense, there's simply a lot of advantage of using 24/196 since the above conversion can result in less loss and higher SNR

as engineers we will never solve this problem as long as the "44.1kHz is good enough" dogma is perpetuated.

here's a question. why are frequency and bit depth the only two variables under discussion here? how does the human ear locate a sound in space? suppose I place a series of 20kHz tone generators along a wall (and that I can still hear 20kHz :) and trigger them at different times, and record the session in stereo at 44.1kHz with a standard X-Y mic setup. will I be able to reconstruct the performance?

So, let’s look at a similar issue with video. Your display is likely only 720p, or 1080p, but a 4K video on youtube will still look a lot better, although technically it should have no visible difference.

But the reality is, we don’t get uncompressed video, or uncompressed audio.

We have a choice between audio or video compressed with lossy codecs, at 16bit/44kHz or 16bit/96kHz, or 4:2:0 video at 1080p or 4:2:0 video at 4K.

And just like you need 4K 4:2:0 mp4 video to get even close to the quality of uncompressed 1080p 4:4:4 video, you also need far higher sampling rate and depth of highly compressed audio to get a quality similar to 16bit/44.1kHz PCM.

That’s the real reason why 24bit/192kHz AAC downloads were even discussed.