KVR Audio

sergiofrias · Post by **sergiofrias** » Mon Apr 11, 2016 10:50 pm

Somethimes we tend to underate synths because they are freeware...
so i compared it against Arturia Sem V :

test A:
(sem v)
https://soundcloud.com/sergiofrias/sem-v
(obxd)
https://soundcloud.com/sergiofrias/obxd

test B:
(sem v)
https://soundcloud.com/sergiofrias/sem-v-2
(obxd)
https://soundcloud.com/sergiofrias/obxd-2

note: obxd is cleaner( less aliasing),in this test i only inverted the polarity on my mixer because the waveforms are inverted and this could bring bad judgement.i also programmed the osc and filter clarity of obxd to better match Sem V
to my suprise they are almost identical,what a great sound indeed.

layzer · Post by **layzer** » Tue Apr 12, 2016 12:43 am

wow. I know 2Dat was passionate about making the osc's and filter as accurate as possible.
funny you should compare it to SEM V, as i started work on a SEM skin for obxd! stay tuned!

zmix · Post by **zmix** » Tue Apr 12, 2016 12:55 am

Arturia is the other company who's oscillators don't alias (much) but they render them with a high noise floor...

aciddose · Post by **aciddose** » Tue Apr 12, 2016 2:04 am

Some properties of a FIR filter are:

1) Slope
2) Stop-band level
3) Pass-band ripple

These are all tied together. If you wanted to maximize slope (to reduce aliasing as much as possible) you'd raise the stop-band level "noise floor" and introduce pass-band ripple.

The aliasing in the stop-band is still aliasing, it's just masked over by noise.

You can use a statistical filter (a special FIR filter) to remove the noise leaving most of the original signal and aliasing behind.

zmix · Post by **zmix** » Tue Apr 12, 2016 2:29 am

That's assuming that the oscillators are filtered. It is possible to limit the harmonics generated, too...

aciddose · Post by **aciddose** » Tue Apr 12, 2016 3:09 am

No it isn't.

It's possible "in theory" or "mathematically speaking" to generate a limited series of partials using a perfect sine function among various other methods.

In reality however it isn't at all possible, not even slightly.

The noise could have some other source, the most likely being part of a wavetable oscillator as it's the most efficient alternative. The most efficient method available however is applying a FIR filter which is why it would be best to assume that was the method used lacking any evidence to better support some other explanation. (Not to mention interpolation used for a wavetable is also generally a FIR filter.)

(If the level is significant it was likely added intentionally at some point.)

If so, the only rational explanation for excess noise is one or both of that it is:
1) Made up of aliased harmonics at the stop-band level of the filter used.
2) Due to quantization error.

In both cases you might argue from a design perspective that adding noise would make sense if noise were measured from the equivalent model oscillator (analog, etc) in order to better emulate that model.

Might... although it would potentially be a bit difficult to come up with any rational explanation for it.

So, alternatively I've guessed that the noise might be intended to mask over the aliases created when synthesizing the waveform. In other words a ruse to get you to say "...doesn't have aliasing, but it has noise". In fact it may very well have both

zmix · Post by **zmix** » Tue Apr 12, 2016 3:28 am

It is possible that the noise was added to mask aliases, yes.

Let's talk about the source of the ailasing.

If the aliasing was generated anywhere "pre filtering" there would be visible sidebands in the FFT, right?

I suspect that the method that is used to generate a waveform at any given pitch at any given host sample rate (i.e. the SRC engine) is the source of aliasing in most soft synths. I have tested them at different host sample rates and noted the differences in the artifacts produced.

aciddose · Post by **aciddose** » Tue Apr 12, 2016 3:43 am

What do you mean "pre filtering" ?

When I'm talking about applying a FIR filter (or interpolation) this is applied to the signal before sampling. So the signal is analog/continuous, we filter it to remove harmonics outside the band and then we sample the result which mirrors the aliases we couldn't filter inside the band.

Any time this sort of sampling happens a filter must be used first to reduce the level of those aliases before they are reflected during sampling.

So in other words to say "pre filtering" doesn't make a lot of sense since the filtering happens first.

aciddose · Post by **aciddose** » Tue Apr 12, 2016 3:46 am

In other words try to imagine "sampling" like placing a box of mirrors around the signal.

The signal originally has harmonics which go on up to infinity although at a level which decreases as frequency increases.

When we "sample" the signal the mirrors go down. Now the harmonics which went up above nyquist get reflected and start going down until zero where they hit the opposite mirror and bounce back up, ad infinitum.

To reduce the level of these harmonics we need to apply the filter first, before sampling takes place.

aciddose · Post by **aciddose** » Tue Apr 12, 2016 3:50 am

zmix wrote:If the aliasing was generated anywhere "pre filtering" there would be visible sidebands in the FFT, right?

Side-bands would be the product of modulation... I'm not certain what you mean by "side-bands" then.

No, assuming we don't apply any modulation/distortion (same thing technically) to the signal at some point further down the signal path there wouldn't be anything present that would be called a side-band.

zmix · Post by **zmix** » Tue Apr 12, 2016 4:27 am

I am talking about aliases, the "foldover" sidebands created by exceeding Nyquist. These can contain frequencies below the fundamental.

aciddose · Post by **aciddose** » Tue Apr 12, 2016 4:31 am

Those aren't "side-bands", that's not the proper use of that term.

They are harmonics of the sampled signal just like harmonics of a ramp/pulse/whatever wave. They're called "aliased harmonics" because the sampled signal is made up of samples which are very narrow pulses (infinitely narrow).

Try with any synthesizer, what is the spectrum of the most narrow pulse you can create? You'll find it's flat with infinite harmonic content. It becomes more flat the more narrow it is, and the amplitude of harmonics decreases relative to the level of the pulse.

(A band-limited version of an infinitely narrow pulse is the sinc() function.)

This approximates what each "sample" is. The harmonics of these pulses are called "aliased harmonics" because when you take a complete sampled signal the combination of these harmonics creates a spectra which is back-to-back mirror images of the band below nyquist.

In other words the 39th image is exactly like the 1st which is exactly like the content within the original band; nomenclature = "aliases". Different names for the same thing.

No, you couldn't see them because they're below a certain level (according to my hypothesis) with noise added on top.

zmix · Post by **zmix** » Tue Apr 12, 2016 4:42 am

I was talking about aliasing, as you recall, but I used the term "sidebands" to describe them because they are generated in exactly the same way as sidebands, they are the product of the harmonics and the sampling rate, same as any modulation that normally produces the artifacts we call sidebands..

Now that we've clarified what we're talking about, lets examine some of the causes.

I've got FFTs of several soft synths producing a 1kHz 50% square wave at 44.1kHz and the aliases are prominent on many commercial softsynths. On the OB-Xd they are nonexistent.

aciddose · Post by **aciddose** » Tue Apr 12, 2016 4:51 am

zmix wrote:they are the product of the harmonics and the sampling rate, same as any modulation that normally produces the artifacts we call sidebands.

No they are not.

The sampling rate is not a signal. If it were, what would it be? A pure partial at the same rate?

In which case the sum + difference side-band products wouldn't produce the same effect.

For example given sr = 20k and signal = 12k:
sum = 32k
difference = -8k

... and yet no spurious harmonic appears at 8k.

aciddose · Post by **aciddose** » Tue Apr 12, 2016 4:51 am

zmix wrote:I've got FFTs of several soft synths producing a 1kHz 50% square wave at 44.1kHz and the aliases are prominent on many commercial softsynths. On the OB-Xd they are nonexistent.

No they are not.

They are simply below the level you have measured or for some other reason invisible to you.

Also just FYI, the proper term is "Fourier transform". You might use "FT" but a better term is "spectrogram" or simply "spectrum". The Fourier transform gives a result in the complex plane or in other words a 2d value for an input signal. You can then compute the magnitude (amplitude) and phase of each "bin" in the transform result and use those to produce a frequency spectrum graph.

Obxd synthesizer