24/192 can capture audio well beyond what we can hear. These sounds can, though, lead to distortion that is within our hearing range, depending on the equipment used.

Therefore, rather than just being useless extra data, it can be actively harmful to the listening experience.

It's not that much different to think of than 480p vs 4K/2160p -- dramatically higher resolution, increasing the width and length of the 'grid' that sound is recorded into -- except that in the case of audio fidelity, quality after a certain point causes problems with playback systems, resulting in audible artifacts and distortion.

That analogy works if you modify it:

- 480p vs 2160p is measuring the resolution of your cell phone propped up on a pillow at the other end of your living room

- experimental evidence shows that your eyesight is not good enough to pick up on the increased resolution, you've maxed out your sensory perception

- Your phone stutters trying to stream at 4k so the playback might actually be worse.

The analogy with images is demonstrated in this screenshot [1], which is referenced in a StackOverflow answer on the topic of image scaling in web browsers [2].

When the high resolution image is downscaled poorly, some high (spatial) frequencies are aliased down into lower (spatial) frequencies, manifesting as blocky/jagged lines. The images are 2D data in the spatial domain, while the audio is 1D data in the temporal domain, but the aliasing of high frequency signal to lower frequency artifact is similar.

Viewing a high resolution image on a panel that has enough pixels to display it without scaling is analogous to listening to 192 kHz audio on a fantastic speaker system that can reproduce those high frequencies accurately, instead of causing distortion by aliasing them to lower frequencies. On the other side, viewing a high resolution image which has been downscaled poorly is analogous to listening to that 192 kHz audio on a realistic speaker system that cannot reproduce high frequencies, which results in those signals aliasing down into the audible range.

And as you say, there is a point where, for the viewer/listener's sake, it doesn't make sense to push for higher frequencies because even if you can build a panel/speaker that will faithfully reproduce those frequencies without aliasing, the eye/ear will not be able to perceive the additional detail.

[1] http://www.maxrev.de/files/2012/08/screenshot_interpolation_...

[2] https://stackoverflow.com/a/11987735

Technically this is not aliasing; rather, the large and varied non-linearities of a speaker can act like a frequency mixer, which is why you'll get a 3 kHz sound (the difference) when playing, say, 20 and 23 kHz.

All good points! Indeed all that extra data makes the decode stage that much more difficult. Same happens with audio too, but high-res audio isn't nearly as bulky as video or nearly as resource-intensive to decode

very good analogy.

Increasing the bits doesn't increase the accuracy of the audio in terms of frequencies. It's 100% accurate regardless of bits. That's why trying to compare it to video resolution doesn't work. Bit depth is more akin to the screen brightness. If you turn down the brightness to 1 step from black, it's impossible to distinguish features in the image.

But weren't those "outside our hearing range sounds that can distort our hearing range" present in real life when the recording was made? So I would have heard the distortions had I been there in real life?

The problem is that your speakers were not designed to reproduce sounds at those frequencies, so instead of faithfully reproducing those sounds (that you can't hear anyway) they may emit unintended artifacts/distortion (that you can hear).

The distortions are not present in the source. They occur because the speaker is a physical device subject to physical limitations and can't perfectly reproduce the waveform. In the process of trying to reproduce the ultrasonic signal, the speaker can produce audible distortions. If the ultrasonic signal is not there to begin with, then it can't possibly cause distortion.

The distortions he's talking about would be introduced during playback by DAC, signal processing, speakers, or amplification.

I doubt that is the case. Most systems which I have used actually sound better at the higher sampling rates. This is not because the sampling rates cause anything to sound better, but because the hardware works better at those rates. In a blind test, you wouldn't be able to decide which one you liked better, only that there is a difference. At least on the machines I have, there are some clearly audible differences.

> you wouldn't be able to decide which one you liked better, only that there is a difference.

Yet people have been reliably _unable_ to do this.

The gold standard in listening tests for this is an ABX where you are simply trying to show that you can discern a difference.

When properly setup and calibrated people are unable to show that they can distinguish 48kHz and 192kHz.

Moreover, by the numbers, audio hardware tends to work less well at higher rates if they're different, because running them at higher rates makes their own internal noise shaping less powerful. (Though for anything done well the distinction should still be inaudible).

Finally, if you do have cruddy converters that sound significantly different at different rates nothing stops you from using a transparent software resampler (SSRC is well respected for good reason) to put the audio at whatever rate you want.. until you get better hardware. :)

"In a blind test, you wouldn't be able to decide which one you liked better, only that there is a difference"

That would have to be the noise. The math doesn't lie...

I'm not talking about noise or other artifacts from the conversion process. I am talking about differences in bass and treble balance that come from different engineering in the converters at different frequencies. In some converters, there are actually completely different circuits that engage for different frequencies. The converters that I have experienced do this, and they all have a noticeable effect on what we hear.

The core is here: https://xiph.org/~xiphmont/demo/neil-young.html#toc_1ch

When you play something stored at 44 khz, there are no ultrasonic sounds recorded. At 192 khz instead, there are, and the speakers may push some of the ultrasonic sounds down to the audible spectrum, causing distortion that the human ear can hear.

It’s mostly because speakers aren’t the best at reproducing ultrasonic frequencies that are captured by the high resolution audio files. You might have components of the speaker that have resonances that manifest the ultrasonic frequencies at a lower frequency.

Imagine the normal operation of a speaker as a swing, except you are pushing and pulling the swing all throughout the cycle as it goes up and down. Now, you can technically move the swing at a variety of frequencies if you’re holding onto it the whole time. However imagine as you push it back and forth (low frequencies), you also vigorously shake the swing at the same time (high frequencies). This would probably result in the chains rattling, similar to the unwanted distortions in the speakers caused by ultrasonic frequencies.

Basically, the sampling theorem says you can reconstruct the exact waveform with a certain number of samples. Adding a bunch more samples bulks up the file, but you didn't need them to restore the exact waveform. However, the unnecessary samples are in the file between you and the next sample you do need. At high enough levels of waste this creates an I/O bottleneck that hampers performance.

Another way to look at is that digital audio is not like digital imaging. There aren't pixels. Increasing the data rate does not continue to make the waveform more and more detailed to human auditory perception in the way that raising the pixel density does for human visual perception.

To describe it intuitively, forget your intuition that audio is like visual and start from "there is no metaphor between these two things."

Digital imaging definitely has sampling concerns too:

https://en.wikipedia.org/wiki/Aliasing

Even analog imaging has concerns that are better described in the frequency domain than the spatial domain:

https://en.wikipedia.org/wiki/Airy_disk

https://en.wikipedia.org/wiki/Optical_transfer_function

Interesting!

Images and audio are pretty similar, and sampling theory works for both - a low-pass filter on audio and a blur filter on images is the same.

The difference is audio is "actually" bandlimited and frequency-based, but images are "actually" spatial. When you try using frequency math on images you get artifacts called ringing, which are easy to see with JPEGs or sharp scalers like Lanczos.

Of course audio isn't really frequency-based either, or else it would just be one sound repeated forever. So there's still ringing artifacts (called "pre-echo") and cymbals are the first to go in an MP3.

> The difference is audio is "actually" bandlimited and frequency-based, but images are "actually" spatial.

I.e. audio is sampled along one dimension, while images are sampled along two dimensions. Note that frequency-domain considerations play a crucial role in all optical design, including imaging sensors.

It's a little more than that. A sine wave originates in the frequency domain but actually sounds like something, but images originating in the frequency domain are just noise and not pictures of anything.

> Another way to look at is that digital audio is not like digital imaging. There aren't pixels.

For the mathematically inclined, this would probably be a good time to repeat: pixels are not little squares[1].

[1] http://alvyray.com/Memos/CG/Microsoft/6_pixel.pdf

> Increasing the data rate does not continue to make the waveform more and more detailed to human auditory perception in the way that raising the pixel density does for human visual perception.

Early low data rate codecs - such as the one used for GSM mobiles - are obviously inferior, but still functional. I think a better analogy is that an iPhone 7 has a 1 megapixel screen, so there's no difference between a 1 megapixel image and a 5 megapixel image, except one is much larger. Of course visually you can zoom in (or move closer in real life), but audibly you can't.

I get what you're saying, that if you remove the ability to zoom you can equalize things, but without that twist I think this preserves the essential problem with the analogy that Monty works so hard to correct. If your data rate is too low, you don't have enough samples to recreate the waveform. So you create an approximate. But there is a finite number of samples that you need to recreate all of the information in the waveform, so once you have that, there isn't any additional information there for you to obtain if you continue to increase the sample rate.

To be fair, phone audio signals have until recently been bandlimited to just a few KHz in centered around 3KHz.

"Diminishing returns"?