Comment by ploxiln | Hacker News Reader

Comment by ploxiln | Hacker News Reader

He compares audio playback quality to food nutrition. It's true, nutrition is far from fully understood. Audio playback as heard by human ears is perhaps not 100% understood, but it's much much closer.

A human being does not need more than 44.1 KHz audio playback. If anyone claims you need a higher sample rate for audio that sounds better, they're lying (whether they realize it or not).

More than 16 bits per sample just lowers the noise floor, and at 16 bits, it's pretty damn low. For playback, you really won't need more. Again, if someone professes the subtle string quartets and powerful metal jams you can only get with more bits per sample, they're lying.

It is true that better headphones will make a real difference, it is true that you can (barely) do noticeably better than 160kbps mp3 (for stereo). But audio engineers understand enough about how the human ear really works to know that more than about 44 KHz or more than about 16 bits per sample truly does not help for final audio playback.

All that to say, this two facts are not like nutrition, and they're not a matter of "some people may hear a difference, but I don't". Selling something as better for playback because of 96KHz / 24bit audio is preying on human weaknesses of judgement, when we know scientifically that these specific two things don't help.

It's been posted a bunch, but this is a really great demonstration of digital audio fundamentals https://wiki.xiph.org/Videos/Digital_Show_and_Tell

A significant factor, which is perhaps not brought up often enough, is the implicit assumption that everything is always going to be 100% linear.

In the textbook case, more than 16 bits per sample might just seem to lower the noise floor. However this is based on the assumption that the voltage spacing is exactly the same for every increment in the sample word -- and this is simply not going to be true in the real world.

This applies to both axes: at 44.1kHz, no clock tick is going to be precisely equally spaced from every other.

In other words, every A/D and D/A process is introducing non-linear behaviour and thus artefacts that change according to the actual input, as a concern quite outside that of the noise floor or the Nyquist limit.

Increasing the bit depth and sample rate helps reduce the effects of such non-linearities.

I really wish people would quit posting Xiph links as if it proves things.

What we DO know is that our ears no not work anyway like how we represent audio in PCM form. We know that they are highly sensitive in some frequency ranges, and that sensitivity includes higher temporal impulse precision than is required, such as the 300-3khz band used for human speech.

As in, for some frequency bands, our "sampling rate" (if such a thing could be applied apples to apples in human hearing) is higher than in others. I suspect this may be related to our ability to localize 3D perfectly sound with only two ears.

Not only that, but also in some frequency bands, we can detect finer differences in sound level than 16 bit can describe.

So no, 16 bit/44.1khz probably is not sufficient.

Also, you bring up lossy encoding which has no place in this discussion. If you want to discuss how lossy codecs exploit the flaws in hearing to compress audio better, talk about codecs that use wavelets and are highly effective at what they do due to a finely tuned psychoacoustic model, such as exploiting that our "sampling rate" and "bit depth" are non-linear across frequency bands.

A reason (probably one of the main ones, IMO) people post Xiph links is because they contain very well reasoned arguments, backed up evidence (i.e the math of how digital audio works, or in the case of the the video linked in the parent, actual meassurement). When seeing arguments on why 16-bit/44.1kHz is insufficient, I more often see faualty reasoning and a lack of evidence to support the arguments being made.

You are correct in that human hearing is highly sensitive in the 300-3kHz band, and that sensitivities to minute timing and phase differences between the ears helps us accurately localize sound with only two ears. But I fail to see how that is relevant to whether or not 16-bit/44.kHz digital audio being enough to reproduce sound without audible distortion.

Are you aware of some evidence that shows how 16-bit/44.1Khz is insufficent? I would be very interested to see it.

There isn't enough research being done on that frontier yet. We know the 300-3khz has a "higher sampling rate" than 44.1khz, but we don't know how high.

However, I'm actually happy with properly mastered 16/44.1 on a proper DAC that has THD and noise levels below what is required for 16 bit (which I can get for $100 in an external DAC, for $150 with multiple inputs, from Schiit).

The big issue is, they don't properly master them, and releasing the HD versions seems to lessen the impact of an improper master (such as we can hear finer steps in volume than 16 bit can do when all the samples have a smaller dynamic range, but we can't if they had mastered everything with 12+ bits of dynamic range like the creators of CDs had intended).

Honestly, if CDs had chosen 20/48 instead of 16/44.1, most of this bullshit would have never happened.

tl;dr I'm a bigger proponent of increased bitdepth over sampling rate, but I'm happy if I get both to get just one.

Edit: Below is probably wrong and I should have actually watched parent's video.

Per Nyquist's Theorem, 44.1 kHz can represent frequencies up to 22050 Hz, which is higher than human hearing's estimated bandwith of 20 Hz - 20 kHz, and for most adults isn't much higher than 17 kHz.

However, take the extreme case of a 22 kHz sine wave, it will be represented in PCM as a triangle wave! So, with 44.1 kHz, the closer you get to 22 kHz, the more distortion / harmonics you get, which may be noticeable.

The point of the linked video is to show that this is not the case. If you plot each sample in the PCM audio on your computer screen and draw straight lines between them it will look like a triangle wave. However, when you play it the sound card will interpolate it back to exactly the same sine wave as you started with. As long as the signal is bandwidth limited to less than the Nyquist limit, sampling and then reconstructing does not introduce distortion.

Interesting, didn't know about the interpolation. Watching the xiph video right now. Out of curiosity, do you have any links that explained how this interp is done?

I'm not an expert, but I think there are two parts to this. First, "oversampling", meaning that in the digital domain one calculates extra samples in between the given ones to make the "staircase" distortions smaller (see https://www.st-andrews.ac.uk/~www_pa/Scots_Guide/iandm/pdf/c...). Second, in the analog domain one applies a low-pass "reconstruction filter", similar to the ones used before sampling. The idea is that the noise added by the staircase effect will be high-frequency, and can be filtered out by removing frequencies above the nyquist rate.

Reply on news.ycombinator.com