48 or 44,1 that is the question

48 for movies or games and 44 for music will do for anyone.
Other sample rates? If you have to ask here then you’ll never notice. And of course you can and should completely ignore us and use whatever Cubase allows you to if you feel it sounds better. We can’t stop you. :mrgreen:

Cryptic answer: 48K

Tiny bit of detail:
() the transition band is nearly twice as large as that of 44K1
() the 44K1 transition band is 1.689 semitones
() the 48K transition band is 3.156 semitones

The reason that few people really care about higher sample rates (that is so say, greater than 48K) :
() Most of what we humans consider to be music happens in a limited frequency range of fundamental frequencies (10Hz - 4KHz)
() Even the 44K1 sample rate captures this limited range with no more than 0.5 dB error per sample
() The frequency region above this range adds important color to the experience but localized / instantaneous phase and amplitude per-sample-errors in this region are far less troubling to most people

peace y’all
pj

Well yes a perfect tranisient has infinite harmonics - it’s basically the transition portion of a square wave. but the intensity of them is low.

I still don’t understand what this means with regard to phase. Are you making the argument that phase is smeared because a transient can start “between” sample points? If the portion of the transient we hear is adequetly reconstructed, why wouldn’t it’s “locality” not also be adequetly reconstituted?

You only need to points pew cycle to reconstruct the wave accurately. Anything laying between those points is of too high a frequency to be sampled anyway. You need an infinite sample rate to accurately reconstruct a square wave, or transient. No one can hear an accurate square wave or transient.

None of this pedantry has any bearing on the enjoyment of the music. :slight_smile:

A single isolated point in a sampled audio stream contains no concrete information about the actual frequency/phase/amplitude of the stream - in fact - even two points can’t characterize the data accurately. It takes quite a few points before a fairly accurate audio stream can be reconstituted into what we know as audio. The good news is this process is well defined for modern sample rates and frequencies below 4K; below this frequency there are enough data points to guarantee that the underlying waveform is “close enough”. By close enough I mean within 0.5 dB as observed in my original post.

I’ve noted this here and elsewhere but perhaps it can be repeated: Nyquist-Shannon requires a waveform which is continuous - the audio signals associated with a track of music are not continuous.

I absolutely agree, Paul

peace y’all
pj

Amen, if it sounds good, do it.

Okay (I know this is pedantry but amuse me for a moment).

According Nyquist, if one takes two samples per cycle of a wave that is bandlimited, then one can perfectly reconstruct the wave (in theory).

The problem as I understand it is, it’s impossible to have a perfectly bandlimited wave… and such a wave would be of infinite duration. Such waves don’t exist in the real world.

But I have seen you at various times pj state that it takes at least 3 data points… and then I recall anothre time where you demonstrated that it actually takes 4 points. I have read several times on the Web that it takes 3. I also recall my geometry classes from High School which taught that it takes at least 3 points to define a curve (plus an underlying equation if I recall) and of course 2 points to define a straight line.

So this difference between Nyquist and what you’ve written confuses me

I haven’t tried higher sample rates but I see there are reasons for running at higher sample rates but for me there are down sides too. If I go from 44.1 to 88.2 I lose use of half my interface inputs, my PC is not old but not up to date either and sometimes is close to it’s cpu limit so I’d have to freeze even more tracks. I’ve read that sample rate conversion can have side effects on the audio.

I guess try it and see is the answer.

Briefly:

Nyquist proposed it in 1928 & Shannon proved it 20 years later – but there is some fine print …
The sampled function must be continuous & differentiable
( And there is a distinction here - note, for example, that the “absolute value” function is continuous but not differentiable: it can be evaluated for all values of input but the slope is undefined at 0 )

The kind of “musical” signals we record are neither continuous nor differentiable
End of Story

yeah, but …

“How come we can record / play back stuff at 44K1 and get pretty good results?”

Magic mostly
And a little math

OK, OK! it’s exactly the opposite - a lot of math and a little magic :wink:

The “What is Best Sample Rate?” question brings forth different emotions and expectations depending on what is really being sought.

If you want a pristine copy of everything in the range of 20Hz-20KHz then 96K is a good starting point though I think 192K is probably slightly better. But I must add that IMO, the only two musical instruments that benefit here are cymbals and “modern” synths. Nearly everything else can be sampled at 48K and survive quite nicely :sunglasses:

If you want to capture a musical “performance” then 44K1 is probably good enough because the information we process as musical “events” takes place at frequencies substantially below that magical 20KHz upper limit of hearing (which is a nominal limit at best - some folks can hear nearly an octave above that and some folks - like me for instance - can’t even hear past 15KHz on a really good day)

peace y’all
pj

What?

Are you saying that some people can hear near 40Khz!!!

Play a note on a just about any musical instrument
Hit a drum - as hard as you like
Scream

All these things have fundamental frequencies which are below 4KHz

I covered this earlier in the thread. At 44K1 samples / second you can acurately track signals of 4KHz and below to 0.5dB which, I think, is good enough for just about anybody.

Modern synths and Cymbals do have significant energy above 4KHz which may be important to some people - and not others.

Hope that helps!

-pj

My first wife could “hear” ultra sonic movement detectors - I believe these are in the 30KHz region
Some times we would walk into a store and she would have to leave almost immediately because she said she couldn’t stand “the sound”. The place was absoluely quiet to me …

Hmm…

I too have heard ultrasonic mouse deterrents, I use a fair few in my country cottage, but what I can hear from a couple of them is lower harmonics or some form of resonance at a audible frequency.

16Khz can be painfull.

lol … I meant what, as in … heh? …speak up … I can’t hear you … cuz your post is like 22khz or something and only dogs can read it.

This sounds far more plausible than someone, particularly as we get older hearing anything beyond 20khz in fact 15khz -16Khz is probably a more typical upper limit for many in their middle years. I’m certainly deaf beyond 16Khz, and I’m sure my -3dB point is a lot lower than that.

But you didn’t directly answer my question Mr PJ. I am sure I’ve read where you wrote that at least 3 points per cycle are needed to reconstruct a signal… and then another time you said four (with reference to coding a vsti like your Nubi if I recall correctly). Likewise, I thought I remember you saying that this 3 data point requirement was one reason why some people say that 60kHz fs is an “optimal” rate

And as I said, I have read a few times other people who have said it takes 3 points. But this confused me because the theorem states only two are required (as Paul pointed out)

Any insight into this would be appreciated

I think perhaps you still miss the significance of the Nyquist-Shannon theorem only “working” for signals which are continuous and differentiable.

I’m sure you’ve read that too; I recall posting ad nauseam on this subject for the past 6 or 7 years …
These days I believe it all depends on the maximum fundamental frequency you will be working with and how close to “perfect” you want the partials above this limit to be captured. My views on perfection haven’t changed - but over the years I’ve realized that the real meat of the musical spectrum is low enough that most of the useful information is preserved even at 44K1.

Yes sir - 60K is a pretty darn cool sample rate - you get 4 samples per cycle at 15KHz and 3 samples per cycle at 20KHz - it is just not a standard rate so I kind’a dig 96K for “everything”

Again, the two point magic happens only for signals which are continuous and differentiable - no free passes on this one :wink:

Next time you are in town I’ll get a pencil and a pad of paper and we can take some time to look at the issues in greater detail and with (I hope) greater clarity

peace to you and yours
pj

Okay, now I get it… funny how it escaped me until now. If you’re sampling at a certain rate, that doesn’t mean you get the same number of data points per cycle on every sine wave. For higher frequencies = shorter wavelengths but since the sampling rate is fixed, you inevitable get fewer data points as you go up in frequency