Repro-1 (out now)

[Urs]

When I change the waveshaper in the resonance path to behave like our Pro-One (self osc goes up by half a semi tone or os)

Wow, I was certain that 3320 uses plain old OTA for reso control based on scarce datasheet info. And judging by saw shape under high-mix-level high-reso condition I would guess that RePro uses something tanh-ish. So, you had to "flatten" reso path transfer function or what, to behave like 3320 when using exact solver?
Have you tried with tanh + large offset? Based on abysmal reso CV feed-trough info in datasheet, if OTA is used for reso it should have lousy specs, up to 10-15 mV of DC offset.

Thing is, when you drive a tanh-ish OTA in a lowpass filter really hard, a sawtooth doesn't magically bend towards a tanh-shape. Because the feedback counters it, it looks more like a slew rate limiter, i.e. the vertical slope of a sawtooth/pulse waveform bends towards a softer slope, until you get a triangle shape for either waveform. In order to achieve a tanh-ish bend you'd need additional diodes in the feedback.

As I said, our Pro-One "goes up in frequency" when increasing resonance on self-oscillation. The only way to do that I could think of is to assume something like linearising diodes that let the slope of the waveshaper in the feedback OTA change in the opposite direction, e.g. have a section of "limit less".

I'm working on some QUCS models to figure this out, but it may be overkill to waste months just on that. Instead we'll use the blackbox approach and just match a bunch of 3320s and surrounding circuitry.

[urosh]

Thing is, when you drive a tanh-ish OTA in a lowpass filter really hard, a sawtooth doesn't magically bend towards a tanh-shape. Because the feedback counters it, it looks more like a slew rate limiter, i.e. the vertical slope of a sawtooth/pulse waveform bends towards a softer slope, until you get a triangle shape for either waveform. In order to achieve a tanh-ish bend you'd need additional diodes in the feedback.

I was referring to characteristic "outward bending" when sawtooth overdrives Pro One alike filter (when feedback element is overdriven negative feedback drops, thus feedforward gain of whole filter increases, so shape of output sawtooth is (crude) approximation of inverse of feedback path transfer function), like on this plot of filter similar to Pro One. Bending in RePro was in ballpark of tanh-alike sigmoid (as far as hardness is concerned), by crude visual examination.

As I said, our Pro-One "goes up in frequency" when increasing resonance on self-oscillation. The only way to do that I could think of is to assume something like linearising diodes that let the slope of the waveshaper in the feedback OTA change in the opposite direction, e.g. have a section of "limit less".

Thing is, I doesn't look like 3320 feedback VCA is OTA+linearisation diodes. Linearisation diodes turn OTA input into low impedance, current input. Datasheet explicitly puts 3.6K resistor at reso VCA input, but even if this is in fact dynamic resistance of diodes, they are woefully under-biased for intended linearisation purpose.

edit: BTW, I got 20 cents drop in selfoscillating frequency pitch with DC offset at OTA-alike feedback element, with feedback gain and DC offset that should be in ballpark of 3320. So, offset it self doesn't explain pitch increase, but pretty please, add reso VCA offset adjustment on final version.

[wagtunes]

I feel like I'm at Gearslutz.

[Urs]

I was referring to characteristic "outward bending" when sawtooth overdrives Pro One alike filter (when feedback element is overdriven negative feedback drops, thus feedforward gain of whole filter increases, so shape of output sawtooth is (crude) approximation of inverse of feedback path transfer function), like on this plot of filter similar to Pro One. Bending in RePro was in ballpark of tanh-alike sigmoid (as far as hardness is concerned), by crude visual examination.

I haven't observed that yet. I don't really want to fry any 3320 just to find that out, and the Pro-One mixer doesn't get so hot

[PAK]

3 sounded the best to me because it produced the less artifacts when modulated at extremes, i.e. it was the best compromise. The difference between 3 and the others was not about 3 bloops instead of 2. It was just night and day..

I chose 3 as the compromise too, but perhaps your definition of night and day is much larger than mine. Maybe that's what happens when you're crepuscular too often

Part of the idea behind this test may have been to see whether people would care for the amount of artifacts vs sound coloration.

Well, in the case of 2 and 5, the colouration was mostly no high end I think some still chose them (especially 5) because they still have some perception issues with the resonance even on 3.

it is an alpha test

Yes. I think that bit bears repeating (even to myself ) Personally, I'll wait for the released product..

And let's be honest, these mods are really extreme.

Yes, and part of that is just how much values shift with a single control movement. I honestly do wonder if this is related to some perceptions since it's not unlike what would be associated with filter stair stepping, except it's caused purely by the huge changes at those extremes, and the limited resolution of on screen controls and regular MIDI.

Like I said, I would totally recommend someone tries to test model 3 with some high resolution control set. It may help certain perceptions, especially if they're related to sweeping the mod knob etc.

[urosh]

I haven't observed that yet. I don't really want to fry any 3320 just to find that out, and the Pro-One mixer doesn't get so hot

I've noticed that on RePro, and figured out that feeback path saturation has to be culprit, don't have ProOne to compare.

BTW, as far as pitch shifting upwards, 3320 datasheet mentions that if first stage is clipped selfoscillation frequency will shift by up to 5%. What's puzzling is their claim that if selfoscillation level is kept bellow clipping, shift should be up to 0.5%, and I have no idea what would cause that 0.5% shift.

[Urs]

I haven't observed that yet. I don't really want to fry any 3320 just to find that out, and the Pro-One mixer doesn't get so hot

I've noticed that on RePro, and figured out that feeback path saturation has to be culprit, don't have ProOne to compare.

BTW, as far as pitch shifting upwards, 3320 datasheet mentions that if first stage is clipped selfoscillation frequency will shift by up to 5%. What's puzzling is their claim that if selfoscillation level is kept bellow clipping, shift should be up to 0.5%, and I have no idea what would cause that 0.5% shift.

Ah, on page 5, yes. I'm sure those 5% are meant to be downwards shifts.

Thing is, once we're done with our Pro-One model, we'll bondage that 3320 (and others) down to a breeding board and measure the individual stages and the resonance cell. I just don't dare to do that before we have a working, plausible model.

[Urs]

Update May 2nd 2016: Our solution & conclusion available in a geeky little PDF:

http://www.u-he.com/downloads/UrsBlog/R ... veiled.pdf

(sorry for the maths stuff, I just can't explain things without that...)

[niektb]

Awesome, you just filled my evening!

[bmrzycki]

Update May 2nd 2016: Our solution & conclusion available in a geeky little PDF:

http://www.u-he.com/downloads/UrsBlog/R ... veiled.pdf

(sorry for the maths stuff, I just can't explain things without that...)

Very interesting read!

One (slightly off topic) question: which of these filter models do you think fit best the Zebra3 development of CPU friendly and mostly analogish sound?

[beely]

Something I've also wanted someone to do, which now seems plausible here:-

What I'd love to see done to discount as much as possible the tactile and physical input to our perceptions, is that when the RePro model is done, u-he get two Pro Ones (preferably one of them non-working ), and:-

Take the non working one, leave the power supply intact, but otherwise hollow it out and put a dedicated computer in there running the RePro model, connected up to the synth's controls.

Now, you have two "Pro Ones", which look, feel, smell, and vibrate with power *identically*.

Take those two synths around and get people to take the "Pepsi Challenge". If the results are statistically different, and people can reliably tell them apart, then when people say "er, no, digital can't do analog sounds like an analog synth" I will take them seriously.

I'm pretty sure I can hypothesise what the results will be...

[EvilDragon]

That was a very good read indeed!

[Urs]

Update May 2nd 2016: Our solution & conclusion available in a geeky little PDF:

http://www.u-he.com/downloads/UrsBlog/R ... veiled.pdf

(sorry for the maths stuff, I just can't explain things without that...)

Very interesting read!

One (slightly off topic) question: which of these filter models do you think fit best the Zebra3 development of CPU friendly and mostly analogish sound?

I don't think there's an easy recipe. We'll have to evaluate this individually. Stereo or Quad filters pose other means of optimisation. They can be parallelised easier, so that this advantage of the expensive algorithms is countered.

Btw. the tricks/cheats we did in Hive and Bazille didn't work at all in RePro. There might have been bugs, but those methods have simply exploded all over the place. They might however work fine for other models, and thus for Zebra.

[beely]

Great, thanks for the update, interesting!

[Urs]

Something I've also wanted someone to do, which now seems plausible here:-

What I'd love to see done to discount as much as possible the tactile and physical input to our perceptions, is that when the RePro model is done, u-he get two Pro Ones (preferably one of them non-working ), and:-

Take the non working one, leave the power supply intact, but otherwise hollow it out and put a dedicated computer in there running the RePro model, connected up to the synth's controls.

Now, you have two "Pro Ones", which look, feel, smell, and vibrate with power *identically*.

Take those two synths around and get people to take the "Pepsi Challenge". If the results are statistically different, and people can reliably tell them apart, then when people say "er, no, digital can't do analog sounds like an analog synth" I will take them seriously.

I'm pretty sure I can hypothesise what the results will be...

Love the idea! (not sure we can do it though)