The aliasing thread

[george]

SFZ 72 quality is really great, but that's not the point here.

The point of the full thread is to discuss about resampling interpolation engines, compare them, post opinions and demonstrate how untrivial is and learn a bit more about it.

And I'm still a bit worried, that some so called HQ-NoAliasing could cutt some frequencies you want from the sound. Resulting into dull sound. But I can't prove that just yet.

Wait a minute... what do you mean? Samplers using poor antialiasing algorythms are better? What about a little research instead automaticly assuming that statement?

By the way, is someone kind enough to tell me what do in order to load a .wav into wusikstation? I couldn't see any guidelines in the 3 page pdf manual

[Majken]

Loading waveforms in W-Station:

Sorry, forgot to reply. Just create a dir called "WAV test" inside the SoundSets folder. And place all WAV files there. That's all, now they show up on the menus. But to be fair, wait 2 days until I finish the Hermite code. I have it ready, it was on Wusikstation, but I had to remove on the last minute. As I have to redo with SSE+Assembly in paralel. (was serial)

There you go, page 4 in this thread

/Majken

[tony tony chopper]

You have to be aware that this web page is a little propaganda for oskari's sampler (that's why rendered results are well hidden).

Antialiasing for this extreme resampling does matter for synthesizers, but not samplers:

-today's soundbanks have a lot of layers. A bank with just 1 sample per octave would be considered as crap. You can even find some with 1 sample per key (in which case it's more disk streaming that matters)

-even the soundfont specs specify that the extreme limit of resampling for SF2 banks should be 2 octaves. Aliasing or not, samples resampled above this will anyway sound very unnatural, so..


For today's samplers, a simple 2x or 4x oversampling would be perfect. That's how I'm gonna do mine when I'll make one. I also prefer to spend the extra CPU, when it's just for little improvements, at rendering time.

[pitchdrifter]

gol, why don't you just for once shut up? Also *if* you say something, get it right, and tell people how you hold oskari's domain name? I guess you'll never get past your totally amateur mindset with anything.
Keep the crap coming man, kvr just wouldnt be the same without you

[george]

You have to be aware that this web page is a little propaganda for oskari's sampler (that's why rendered results are well hidden).

Jeskola is actually the fourth sampler in the lists, and the stated things there are correct, unless you can prove I'm mistaken, of course.

A bank with just 1 sample per octave would be considered as crap.

Sure, like the hundreds of AKAI, EMU or Kurzweil sample CDs. All crap Still, a poor resampling will result in plenty of "musically unnoticeable" digital junk even when a semitone up or down is used.

For today's samplers, a simple 2x or 4x oversampling would be perfect. That's how I'm gonna do mine when I'll make one. I also prefer to spend the extra CPU, when it's just for little improvements, at rendering time.

What about using simple 2x or 4x oversampling in fruity sampler too? Live musicians will benefit of that for sure

[george]

gol, why don't you just for once shut up? Also *if* you say something, get it right, and tell people how you hold oskari's domain name? I guess you'll never get past your totally amateur mindset with anything.
Keep the crap coming man, kvr just wouldnt be the same without you

Please, take it easy. We are not here to discuss about ethics but resampling qualities.

[george]

Sorry, forgot to reply. Just create a dir called "WAV test" inside the SoundSets folder. And place all WAV files there. That's all, now they show up on the menus. But to be fair, wait 2 days until I finish the Hermite code. I have it ready, it was on Wusikstation, but I had to remove on the last minute. As I have to redo with SSE+Assembly in paralel. (was serial)

My bad, I overlooked your answer, sorry. I will wait until wusikstation update then.

[tony tony chopper]

Jeskola is actually the fourth sampler in the lists, and the stated things there are correct, unless you can prove I'm mistaken, of course.

It's just that I remember I mailed to about showing rendered results as well, and you were seeing them as not very important.

What about using simple 2x or 4x oversampling in fruity sampler too? Live musicians will benefit of that for sure

I did that in sytrus. FL's sampler channels would benefit from that, but I opted for low CPU consumption instead. You have (but you know that since you made the test) plenty of quality options at rendering time.

I was just saying that, if I had to make a sampler today (and I'll make one in the future), I'd choose a user-selectable oversampling, as in Sytrus. Mip-mapping is a cool solution, but I'm lead to believe it can f**k up looped regions, and if a 4x oversampling [perfect for 1 sample per 2 octaves] is ok, CPU-wise it would be ok as well.
Aliasing can happen in other places anyway, like distortion - which sampler doesn't have its built-in effects today?

a poor resampling will result in plenty of "musically unnoticeable" digital junk even when a semitone up or down is used

for that other aliasing, a common cubic interpolator seems to do very well. My future sampler will have its default quality set to 2x oversampling and cubic interpolation. Or maybe a 6point hermite since I already have the optimized code. But I wouldn't even go for sinc anymore.

[tony tony chopper]

Also *if* you say something, get it right, and tell people how you hold oskari's domain name?

oh, and we hold oskari's domain name hostage

(for those who are interested)

[Majken]

I tried it on my girlfriend to and asked if she could spot any overtones above the base notes. She could only hear a difference on the lowest descending notes but said "the clip called Wstation has em and the one called SFZ 72 doesn't"

Gol: Yes, alot of sample libraries are using multiple samples and most pitch mless than an octave. However in the W-station example i can hear aliasing in the 2nd and 3rd seminote pitched downwards. Now from what i understand w-station may not be the best at low aliasing but it is definately not the worst either. Meaning any AKAI library could have obvious aliasing in the wrong sampler. Which is why i find this discussion highly interesting.

/Majken

[stefancrs]

[snip]

they produce diference results with a 15Khz sine pitched 4 octaves up. But what I'm trying to prove is that in "normal" usage, they will all sound pretty much the same. And I'm still a bit worried, that some so called HQ-NoAliasing could cutt some frequencies you want from the sound. Resulting into dull sound. But I can't prove that just yet.

Who is talking about pitching a 15KHz sine 4 octaves up? (not that it matters, really good filtering actually would handle that, the result would be silence if using for instance 48KHz sample rate).

And about cutting frequencies you _want_ in the sound. The whole idea with antialiasing is cutting everything _above_ nyquist, why would you cut anything else? It'd be a really crappy method you'd be using then.

Thing is, most high quality samples contain stuff all the way up to nyquist, if you pitch these up anything at all without filtering out the stuff above nyquist, you will get aliasing. If the spectral content in the high frequencies have a high amplitude, you will get very audible aliasing. You should never assume that the musician only will use samples without high frequency components.

[george]

It's just that I remember I mailed to about showing rendered results as well, and you were seeing them as not very important.

I dont think so, since your mail made me consider to put rendered results too, in order to settle your dissapointments.

I'd choose a user-selectable oversampling, as in Sytrus.

Wise decition.

Aliasing can happen in other places anyway, like distortion - which sampler doesn't have its built-in effects today?

It happens but can be easily avoided by using oversampling, perhaps something to consider as a high quality rendering option.

for that other aliasing, a common cubic interpolator seems to do very well.

It does indeed, but if we look at how resampling qualities are being developed nowadays, making use of cubic interpolation only in new products is a step backwards.

My future sampler will have its default quality set to 2x oversampling and cubic interpolation. Or maybe a 6point hermite since I already have the optimized code. But I wouldn't even go for sinc anymore.

Hermite seems quite ok for realtime purposes, however Jeskola XS-1, after almost two years still has got the best resampling quality:cpu usage ratio.

[tony tony chopper]

However in the W-station example i can hear aliasing in the 2nd and 3rd seminote pitched downwards

yes but that's not the same kind of aliasing. For this aliasing, no interpolation is very noisy, linear is a big improvement but still a little noisy, and cubic/hermite/whatever with 4 or 6 points is really good enough.

This aliasing is like when you draw a line with or without aliasing, to smooth the jagged edges. The other aliasing is more a 'moirage'.

[tony tony chopper]

however Jeskola XS-1, after almost two years still has got the best resampling quality:cpu usage ratio yet

how about quality:cpu:ram?

It happens but can be easily avoided by using oversampling

..which you can use to reduce resampling aliasing as well

It does indeed, but if we look at how resampling qualities are being developed nowadays, making use of cubic interpolation only in new products is a step backwards.

what are the other options anyway? You're limited to 2 to 8 points if it has to run realtime. It's not like you were gonna use 500 points and keep this realtime.
Combining a 4 or 6 point hermite AND oversampling seems a better solution these days.

[george]

how about quality:cpu:ram?

I can't tell anything about it since I havent made any researchs, sorry.

what are the other options anyway? You're limited to 2 to 8 points if it has to run realtime. It's not like you were gonna use 500 points and keep this realtime.
Combining a 4 or 6 point hermite AND oversampling seems a better solution these days.

If I'm not mistaken, SB Live uses 8 point SINC, and results _may_ be better than hermite + 2x oversampling in CPU and quality terms.