Any use of going beyond 44100?

[pitchdrifter]

Bunch of armchair travelling VIPs around here me thinks.

[DevonB]

Ok, you tell me I'm wrong, but won't tell me why? Try again.

You are wrong and I did explain why. Contrast then gave another explaination why you are wrong. I'll try again if you like?

Please, do so. I'm referring specifically to how 44,100 samples taken in 1 second time interval of the amplitude of the signal is NOT playing connect the dots.

Devon

[pj geerlings]

It all comes down to this: higher sample rates have much broader transition bands.

The transition band is the frequency band between the highest frequency you care about (typically 20KHz) and the actual Nyquist-Shannon frequency that is always half of the sample rate.

In the case of a 44.1K sample rate the transition band may be specified as 22.5K/20K which equals 1.125 – in log base 2 terms this value works out to 0.17 of an octave. A real-world filter that works over a transition band this small is very difficult to design. (These days, oversampling techniques are used in the DAC hardware to relax the filter requirements for any sampling rate – but for the moment lets not consider this aspect of the issue)

If we bump up the sample rate to 48K things get better. 24K/20K equals 1.2 – or 0.26 of an octave. Designing a filter to work over this transition band is not nearly as difficult. In fact, this seemingly tiny bump in sampling rate reduces the required filtering complexity by an order of approximately 2 to 1.

Even if you factor in the hardware assisted oversampling, for any given sample rate the higher the sample rate the broader the transition band will be and that means a better realization of the signal in the real world.

Jumping up to 88.2K or 96K sample rates makes filtering almost trivial. But at a cost: these rates do double the computational complexity of the entire process and some CPUs are just not up to the task just yet.

A final note: there are still people in the world who use magnetic tape to record audio. The bias frequency of the really high-end machines is approximately 400KHz – while it is not exactly equivalent, intuitively this translates into a very high sampling frequency. At 192K samples per second we are still only half of the way to this goal.

[stefancrs]

Disagree here. Higher sample rate will give more accurate recreations, and on some VST's, yes it does make the sound a bit smoother. You're playing connect the dots with 96,000 points or 44,100 points over a one second segment. You tell me which would be more accurate and smoother? Also, 96k is going to prevent some aliasing too, at least for freqeuncies up to 48kHz instead of only 22.1kHz.

Devon

When "connecting the dots" (like you do when recording digitally and then playing back) sample rates above twice the highest perceived frequency won't do you any good whatsoever.
On VSTi's (and VST's) it might matter, but that's not because of the reason you stated. It certainly has nothing to do with accuracy. When running a aliasing generator at 96KHz it will still produce aliasing, and this should be compared to running a alias free generator (which most of us prefer) which won't alias even at 500hz sampling rate.

When the content is created in the digital domain it's no longer about connecting the dots, and when we're speaking about "recreations", the best sample rate is dependant on your adc (if you yourself recorded it) and most of all, the dac. As long as you're above 40KHz that is.

[stefancrs]

Ok, you tell me I'm wrong, but won't tell me why? Try again.

You are wrong and I did explain why. Contrast then gave another explaination why you are wrong. I'll try again if you like?

Please, do so. I'm referring specifically to how 44,100 samples taken in 1 second time interval of the amplitude of the signal is NOT playing connect the dots.

Devon

It is connecting the dots when you convert it from the digital domain. Regardless of samplerate you can still recreate any signal below nyquist perfectly.

[stefancrs]

[lots of interesting stuff]

Hm, why do you have to filter in the "transition band"? Are you speaking of ADC'ing or DAC'ing?

[bmanic]

Bunch of armchair travelling VIPs around here me thinks.

Indeed.. These are the kind of threads that remind me that this is KvR.

[Borogove]

Please, do so. I'm referring specifically to how 44,100 samples taken in 1 second time interval of the amplitude of the signal is NOT playing connect the dots.

Connecting the dots looks intuitively like a good idea. Connecting the dots is a good approximation when the signal you're sampling is very low frequency compared to the sampling rate. But connecting the dots isn't Correct.

Consider the case where you're sampling a full range sine wave at exactly 1/2 the Nyquist rate (= 1/4 the sample rate, = 11025 Hz for 44100Hz sample rate). If the sampling phase exactly matches up with the phase of the sine, then you'll get the following samples out:

0, +1, 0, -1, 0, +1, 0, -1...

If you "connect the dots", you'll see a triangle wave, but this isn't a triangle wave - it can't be a triangle wave, because a triangle wave at that frequency would be carrying substantial energy well above the Nyquist rate, which we know from sampling theory is impossible. Hence reconstructing the analog signal corresponding to that string of samples must be something other than "connecting the dots".

(Another example worth considering is a sine wave very close to Nyquist. If you sample that and then connect the dots to reconstruct it, what you get out looks like an amplitude-modulated triangle wave. The amplitude modulation also implies energy above Nyquist, and thus is wrong.)

The correct way to go from the sampled representation to the perfect analog reconstruction is to go through a lowpass filter with a very very steep cutoff - as close to infinite slope as you can get. Because a sharp cutoff in frequency-space implies a long impulse response in the sampling domain, it turns out that every point you've sampled winds up affecting the reconstructed curve at every point in the curve, even sampled points very far away from the portion of the curve you're considering. (Look up 'sinc reconstruction' on google if you're ready to bake your noodle.) Fortunately, the influence of each point does decrease with distance, so at some point you can stop worrying about the influence of far-away sample points.

[_starcraft_]

i thought it was all nonsense before trying it.....but indeed SOME vstis do sound alot better at 96hz.
wether this is due to poor code or not i have no knowledge to judge......but the sound at 96hz is clearly fuller.
it would solve most problems if all vsti supported oversampling.
for everything else 44.1 hz seems to be just fine.

[stefancrs]

i thought it was all nonsense before trying it.....but indeed SOME vstis do sound alot better at 96hz.
wether this is due to poor code or not i have no knowledge to judge......but the sound at 96hz is clearly fuller.
it would solve most problems if all vsti supported oversampling.
for everything else 44.1 hz seems to be just fine.

now go try some distortion fx...

[pj geerlings]

[lots of interesting stuff]

Hm, why do you have to filter in the "transition band"? Are you speaking of ADC'ing or DAC'ing?

Actually both

The existance of a transition band is what destroys any hope of the theoretic "perfect" reproduction (and recording).

The filter that is required (both in to and out of a digital system) must pass all frequencies of interest with no attenuation and completely attenuate all frequencies above the Nyquist-Shannon limit frequency. A transition band always exists for any sample rate because real-world filters will never have infinite cut-off rates.

The existance of a transition band seems to be the most over-looked aspect in these discussions. IMO, understanding it is very important in understanding the limits of "perfect" digital reproduction.

peace,
pj

[bmanic]

yeah, distortion is one of those that always seems to benefit (unless you want it to sound gritty) and same goes for 99% of the EQ plugins out there.

EDIT (now corrected spelling too)

I forgot to add that a lot of old VSTi's might sound "worse" because the filter changes it's cutoff frequency and Envelope/LFO behaviour can also change when ran at 96khz. Just re-adjust the patch to sound identical to the original and all should be fine.

- bManic

[kuniklo]

i thought it was all nonsense before trying it.....but indeed SOME vstis do sound alot better at 96hz.
wether this is due to poor code or not i have no knowledge to judge......but the sound at 96hz is clearly fuller.
it would solve most problems if all vsti supported oversampling.
for everything else 44.1 hz seems to be just fine.

Exactly. Some do, some don't. I suggest anyone curious about this bump their sample rate up to 88k or 96k and audition a few of their favorites. If you can't hear a difference then there's no point. I was surprised myself.

z3ta is also an interesting test for this. Some patches sound identical at 2x oversampling. Some sound quite different.

[stefancrs]

[lots of interesting stuff]

Hm, why do you have to filter in the "transition band"? Are you speaking of ADC'ing or DAC'ing?

Actually both

[the well carried explanation here]

wtf, I knew that this trying-to-stop-smoking is seriously damaging my brain, short-term.

[DevonB]

Please, do so. I'm referring specifically to how 44,100 samples taken in 1 second time interval of the amplitude of the signal is NOT playing connect the dots.

Connecting the dots looks intuitively like a good idea. Connecting the dots is a good approximation when the signal you're sampling is very low frequency compared to the sampling rate. But connecting the dots isn't Correct.

Consider the case where you're sampling a full range sine wave at exactly 1/2 the Nyquist rate (= 1/4 the sample rate, = 11025 Hz for 44100Hz sample rate). If the sampling phase exactly matches up with the phase of the sine, then you'll get the following samples out:

0, +1, 0, -1, 0, +1, 0, -1...

If you "connect the dots", you'll see a triangle wave, but this isn't a triangle wave - it can't be a triangle wave, because a triangle wave at that frequency would be carrying substantial energy well above the Nyquist rate, which we know from sampling theory is impossible. Hence reconstructing the analog signal corresponding to that string of samples must be something other than "connecting the dots".

(Another example worth considering is a sine wave very close to Nyquist. If you sample that and then connect the dots to reconstruct it, what you get out looks like an amplitude-modulated triangle wave. The amplitude modulation also implies energy above Nyquist, and thus is wrong.)

The correct way to go from the sampled representation to the perfect analog reconstruction is to go through a lowpass filter with a very very steep cutoff - as close to infinite slope as you can get. Because a sharp cutoff in frequency-space implies a long impulse response in the sampling domain, it turns out that every point you've sampled winds up affecting the reconstructed curve at every point in the curve, even sampled points very far away from the portion of the curve you're considering. (Look up 'sinc reconstruction' on google if you're ready to bake your noodle.) Fortunately, the influence of each point does decrease with distance, so at some point you can stop worrying about the influence of far-away sample points.

Thank you. Seems I've been grossly oversimplifing the process.

Devon