Comment by louthy | Hacker News Reader

Comment by louthy | Hacker News Reader

Parent Full thread

> Not true at all..

Not true at all.. the bit rate will only affect the noise floor, you'd be hard pressed to notice any difference between 16 and 24 bits (you won't). And when it comes to the Hz - (width of the time-slice, 44 / 96) that will only make any difference on super high frequency sounds, but again, even with that the vast majority of people will perceive no difference. Band limited sound waves can only solve to one solution, and making the slices smaller in the time domain will have relatively little effect because you can perfectly represent all signals in the range of human hearing at 44khz sample rate.

Higher bit rates are useful for people using DAWs like Cubase, ProTools, etc. where your 'signal' is going through myriad processing and summing stages because of the use of multiple tracks, plugins, etc. (sometimes hundreds) That's when you don't want the artifacting from summing multiple approximations. I use 32bit / 96khz in Cubase for this reason.

The human brain is very good at convincing itself [3] that it can tell the difference, but in a double blind test you'd have no chance.

Useful info and reference:

[1] https://www.youtube.com/watch?v=cIQ9IXSUzuM

[2] https://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampli...

[3] http://www.image-line.com/support/FLHelp/html/app_audio.htm

So you say 'not true at all' to my repudiation, and then go on to say:

"you'd be hard pressed" and "the vast majority will percieve no difference"

which is it?

bit rate: yes, as sounds are at the lower end of amplitude within the bit depth, the more 'stepping'/'zippering' occurs to the input waveform since less bits are available to represent the shape of the wave at that amplitude, and so things become more distorted, and the signal noise goes up. similarly, for higher amplitude sounds, the less need for compressing/limiting to remove clipping artifacts, so again, less distortion (though the latter isn't inherent to the signal encoding per-se).

sampling frequency: irrespective of the nyquist theorem, the shape of the wave is different. even if you can't explicitly hear at a base frequency of the overtones above a certain frequency, the waveform is not the same and your ear is not being moved in the same way, period, and relative distortion of overtones with respect to the base frequencies within the nyquist spectrum increases with frequency, so the tonal effect of the sampling frequency on the sound (and net distortion) is diminished as the sampling frequency increases. aka overtones become increasingly distorted to a lesser extent as the sampling frequency increases; banded spectrum (e.g. FFT display) is much cleaner, as a visual example.

both of these can be masked somewhat with dithering, but again, this is introducing noise into the signal, which you can also hear; all other things equal (which they arent), the dither free signal will inherently have less noise and so be cleaner sounding.

I agree that bit rate is harder to perceive, and for most things, this is more important w/r/t multiple generations of distortion introduced by software.. but just because something is 'hard' to percieve doesn't mean it is imperceptible and therefore a net zero, which is what the original post is claiming. whether a net 0.0001 is relavent is a different matter (as I said previously, 16/44 is generally fine, so I'm not in the 'I can totally hear the difference from my monster cable digital edition' camp), but it is not a net 0, and when listening for tambre rather than frequency, quite clear once you know what to listen for.

also yes:

I did a blind listening test for this while in an audio engineering program. Was able to clearly tell 44 vs 96 and vs 192, as well as 16 vs 20 and 24 (though 24 vs 20 was pretty difficult). That was a while ago so not sure if my ears are still as sharp.

How about percussive sounds, such as a cymbal crash? These show very high frequencies in a Fourier analysis. Wouldn't that invalidate the application of the Nyquist-Shannon theorem?

It's worth watching the video in my references [1] - it's a proper hacker's view of digital audio and how it translates to the analogue domain. Even with a 20khz sound you're not going to get any aliasing at 44khz sample rate. Most people's hearing starts to fall off significantly after around 12khz or so (depending on age). Most cymbals have a fundamental frequency around 2-5khz in my experience.

Also, because cymbal sounds are pretty close to being noise, it's arguable whether it would matter or not. Although if you listen to a 128 kbps MP3 then you can often hear the top end frequency sounds (like cymbals) becoming 'mushy'. To the vast majority of people once you get to 192 kbps it's inaudible, and then above that most people with 'perfect ears' would struggle. And that's with a lossy format, never mind the lossless formats of CDs, WAV or FLAC which is the discussion at hand.

[1] https://www.youtube.com/watch?v=cIQ9IXSUzuM

The ADC (Analog-digital-converter) will always have an anti-aliasing lowpass filter to remove any frequencies above the nyquist and prevent aliasing in the recording. Sometimes it's analog, nowadays it's often a 2-stage system with oversampling and a digital anti-aliasing filter. But the result that goes into the recorder is the same, any frequencies above the Nyquist (which would become aliased and mirrored) are removed or reduced to inaudible levels before being finally digitized.

So while cymbals do have lots of high frequency content, in a 44.1khz recording any frequencies at or above 22.05khz should be removed first. (In reality it's not a brickwall filter at that frequency, but tends to steeply roll off once it gets nearer).

Unless I'm misunderstanding what you're asking.

Reply on news.ycombinator.com