The aliasing thread

[tony tony chopper]

There is no bad faith. The subject is about real time and if your product doesn't excel in that specific area, I'm not going to hide it, although fruity offline rendering is top notch, so why bother?

Ok, then define realtime. This kind of stupid test is what make companies add stupid features, just to look better in a test.

I'll probably end up allowing the user to switch the sinc interpolator realtime in FL. Result for the user? None, he will be able to play 2 notes, so will switch it off. Result for us? For the product to look better in a stupid page.

A good test would show quality vs CPU vs ram in a chart. You'd then let the user switch what matters the most to them. But 'realtime' is stupid - it doesn't compare CPU at all. You could have some good interpolators already listed in your 'realtime category', yet they might just eat too much CPU in a complex project, so not be that much realtime.

If you want to promote your tests, just do them right. It's not that hard to check the CPU usage of something. It's not hard to understand that everything is a CPU vs quality vs memory usage tradeoff, and that all 3 parameters matter.

[Sicklecell666]

So you can't discuss this topic properly and start harrasing me instead? If you can't keep this thing civil I'd ask you to move on.

I will do exactly that. I apologize if it appears I am questioning your professional integrity, but I have seen enough ugliness from you in the past I difficulties digesting anything you say.

With respect to you as a developer, I will leave my personal feelings out of the discussion & simply avoid this thread.

I hope ALL developers involved in this discussion are able to benefit from it & implement what's been said here into making better instruments. That doesn't mean I have a compliant about ANY of you, it just means as an Artist myself, I am aquainted with the never-ending & un-attainable stretch for perfection we all yearn for..

later, George.

I wish I could get past how I feel, but I doubt it matters to you anyways.

[René]

Remember that I did the same mistake with another program and someone jumped on me like a wolf. (no one never said sorry to me here so far)

Yes. And I doubt anyone will tell you sorry. If you revise the thread, gol DID ASK ME to test his interpolation routines. I wouldn't ask you to measure the area of a circle, so your test was unrequested.

-René

[experimental.crow]

gol wrote ...

... It's not hard to understand that everything is a CPU vs quality vs memory usage tradeoff, and that all 3 parameters matter.

this is the crux of the biscuit ...
what can i , joe consumer , expect on my average machine ?...

[tony tony chopper]

gol DID ASK ME to test his interpolation routines

no, you were claiming that downsampling introduced a noticable latency, and I asked you to test Sytrus to check that it wasn't the case. To which you replied that you didn't see any point, since you couldn't load samples nor check noise(?)
Couldn't you just trust me that it does interpolate down and check that there's no latency added? If you don't, then make your own tests, but don't claim it can't work.

[schlsdwow]

i'm very surprised that there are developers making plugins who are either a) not checking the output in an analyzer or b) don't care what that shows

surely you are not using only your ears to figure out what your code is doing?

the only reasonable way to see what effects your coding is having on the sound, is to use some analysis. Why are analyzers suddenly 'useless fancy graphics'? What reasonable developer or user would take this stance, and would you trust them?

and why are consumers of this technology vocal in their refusal to accept what graphic analysis shows? just use some freeware plugin if you don't care about the quality.

seriously, i'm surprised at so many things in this thread, yet grateful at the efforts of some developers to try not only create the best quality instruments, but also to try to show other devs how to achieve better results.

by why the outrage from users? we are all winners when devs make the choice to measure and benchmark their products, and strive for better

geez

[tony tony chopper]

'useless fancy graphics'

the graphics are useful, and I believe (hope) show the truth. Only problem is that.. they're biased. Lots of parameters are not taken into account. True, it doesn't claim to hunt for the best sampler, but the best resampling methods in samplers. But even here, some parameters are neglected. The best resampling method is not ONLY the one that gets rid of aliasing.

[René]

no, you were claiming that downsampling introduced a noticable latency, and I asked you to test Sytrus to check that it wasn't the case. To which you replied that you didn't see any point, since you couldn't load samples nor check noise(?)
Couldn't you just trust me that it does interpolate down and check that there's no latency added? If you don't, then make your own tests, but don't claim it can't work.

I didn't say it can't work.
I did my own tests for Sytrus, check the url in my previous post.

Even when it really surprises me that Sytrus alias the two last notes at 64x oversampling (which a typical 2x oversampling does not), I think it's pretty good. Congratulations.

-René

[tony tony chopper]

Even when it really surprises me that Sytrus alias the two last notes at 64x oversampling

It might be, it just depends on the source oscillator (which you can pre-filter using a slider).

This is because it's a synth, so the sampling rate of the source samples is in theory infinite, and in practice very big, so contains very high freqs (in a way).
For a sampler (and yes I tested with small samples), it'd be the same except that the samples have a much more finite & realistic samplerate, that's around the pitches of the keyzone they're in. So they already don't contain any freq above their root freq(their own nyquist).


(also for the record, I know there are much better way to antialias an oscillator, but there's no other choice than oversampling for synths, like Sytrus, that feature free shapes, FM, RM and waveshaping. So in a way, if a sampler features waveshaping (many samplers provide a built-in distortion fx these days), it'd need to do a bit of oversampling, and this oversampling could be as well part of the sample's own resampling)

[george]

I do think that it would be really cool to include the results of samplers which can deliver a better interpolation on rendering.

It is being covered too, there is a link to offline renders sample list. I'm afraid I didn't know about that sfz mode.

I'll probably end up allowing the user to switch the sinc interpolator realtime in FL. Result for the user? None, he will be able to play 2 notes, so will switch it off. Result for us? For the product to look better in a stupid page.

I agree, being that extreme gives no results. Showing off some flexibility using reasonable algorythms like hermite or sinc16 will do.

For the product to look better in a stupid page.

If you think it's so stupid, why argue about the subject anymore? Just keep customer base happy with linear interpolation, because they still can make high quality renders.

A good test would show quality vs CPU vs ram in a chart. You'd then let the user switch what matters the most to them. But 'realtime' is stupid - it doesn't compare CPU at all. You could have some good interpolators already listed in your 'realtime category', yet they might just eat too much CPU in a complex project, so not be that much realtime.

If you need a reference, I can tell you realtime is being considered as something who can provide at least a single note poliphony on a p3 1.1Ghz CPU Still, not a real life example, tho.

If you want to promote your tests, just do them right.

I have said something about blaming before. Test are right, and linear interpolations go at the bottom of the comparison not because I want but because there are competing products doing this task much better.

[Muon Software Ltd]

Hang on. There's more going on in that spectrogram than just aliasing of the source signal.

If I have a sinewave at 15khz, and I pitch it up one semitone, it isn't going to go over Nyquist at 44.1 khz sample rate - therefore it isn't going to alias in the strictest sense of the word (one semitone up from 15khz is 15891.94642khz)

To go over Nyquist @ 44.1khz, you'd need to pitch the 15khz sine more than 6 semitones (7 semitones shift=22474.60615khz). Then you'd expect to see an image at 21625.39385khz - and since it is a sine wave, you'd expect to see only ONE image on the spectrogram.

So in fact, those spectral images around the source wave perhaps shouldn't strictly be referred to as "aliasing". They are in fact the byproducts of an interpolation algorithm that doesn't take the frequency domain into account. Linear interpolation, hermite interpolation etc. all are strictly join-the-dots types of interpolation that look OK in the time domain but have non-linear effects rather like a waveshaper in the frequency domain - they can synthesise harmonics that really shouldn't be there!

IMHO what the spectrogram shows is not aliasing in the strictest sense, but distortion resulting from a non-linear process such as a computationally cheap non-bandlimited interpolation algorithm. It should be noted that the distortion generated by a "naive" time domain operation like this can alias too.

Now, people know that I know my aliasing but to be honest, in a digital *sampler*, I probably couldn't care less. In an analogue synth emulation I'd be the first to jump in and say no, there should not be aliasing in the oscillators - analogue oscillators do not alias, full stop end of story.

However, in a digital synth or sample, or a non-emulation synth, does it really matter? my Yamaha S30 has horrible aliasing. So does my Akai S3000XL. I never really checked before I sold the JV2080 but I bet it aliases to hell too.

So I sort of have to wonder if all of this is just p155ing into the wind. I would be happy to have DS404 near to the bottom of the chart because it only uses linear interpolation for speed and I guess it doesn't sound all that bad

Regards
Dave

[tony tony chopper]

why argue about the subject anymore? Just keep customer base happy with linear interpolation, because they still can make high quality renders

Just because someone can jump to your page, see what you show like a chart of sampler qualities, and judge them by this.
Just show them all of the parameters, and everything they should take into consideration. Maybe some would be interested to check the CPU usages as well, because it matters? (well if CPU doesn't matter, I'll try to fit some more features in Sytrus [which is realtime], that's already considered as eating a bit too much)

[autloc]

Linear interpolation, hermite interpolation etc. all are strictly join-the-dots types of interpolation that look OK in the time domain but have non-linear effects rather like a waveshaper in the frequency domain - they can synthesise harmonics that really shouldn't be there!

Hmm, maybe too tired and should wait to sober up completely -- maybe misunderstanding what you are saying -- but surely linear interpolation (!), hermite interpolation and the rest are all effectively FIR filters (i.e., linear operations) - they're just of different order and using different methods to select the FIR coefficients. An FIR filter (well, any linear operation in the time domain) cannot add spectral information, only boost or attenuate, so it is the act of changing the sample playback rate (i.e., resampling) that is causing the distortion, not the application of an FIR filter.

[george]

Just because someone can jump to your page, see what you show like a chart of sampler qualities, and judge them by this.

gol, the page states there are offline renderings, available in the same line of the product image link.

I do believe you are assuming something that may not happen at all. I can isolate all the products using offline rendering links into another section, if that makes you feel any better

[Sascha Franck]

FWIW, I just did the 15kHz test with both Kontakt and the EXS... indeed horrible results!
And then - while I was thinking all the test was just some leet coder BS which wouldn't get in my way, it defenitely isn't. Even someone with my "way less than golden ears" is easily able to listen to the dramatic effects on some of the notes.

And while you usually wouldn't sample a 15kHz sine, such frequencies easily are part of the harmonics in other samples.

I allready noticed the rather bad behaviour of the EXS when sampling some stuff without using a lot of keymapping. Back then when the EXS came out I was comparing some "1 sample for the whole mapping" patches in both the EXS and my (shitty) EWS64XL soundcard's onboard sampler. Quite a lot of those patches came out as being totally useless in the EXS while the EWS performed halfway fine (so I guess it wouldn't do too bad in these tests, can't test anymore as it's ISA).

Allright, what to do about it?
When I was running into those disturbing noises for the first time (I had no clue what they were caused by back then) I was rather frustrated... but then, these days I just do some better keymappings.
Quite a deal sometimes - I'd really prefer if I'd still be able to do some single sample patches. It just doesn't seem to be possible with Kontakt or the EXS right now.

Anyways, I do now strongly believe that those tests are no scientific bullshit. The (particularly bad) results do indeed have a more or less strong impact on your overall sound.
Maybe one day I just want a sampler doing things better than my current tools of choice (any chance you'll be doing a real sampler, Rene? )