Vowels

[martinvicanek]

Another nice feature about the decomposition into five 2-pole filters is that you can morph any two (or more) vowels by linear interpolation of the filter coefficients. The filter is stable by design. Attached is a human sounding wah-wah demo.

[tester]

But... there are more vowels.
http://en.wikipedia.org/wiki/Vowel_diagram

or to emphasize the understanding through google search:
http://www.google.pl/search?q=vowel+cha ... 18&bih=676

I mention this, because my language speaks differently english vowels.

[MegaHurtz]

I have wondered a couple of times which filters would be best to do this with, it seems linear phase isnt quite it.

[stw]

Another nice feature about the decomposition into five 2-pole filters is that you can morph any two (or more) vowels by linear interpolation of the filter coefficients. The filter is stable by design. Attached is a human sounding wah-wah demo.

Nice work Martin! Thanks for sharing...

[martinvicanek]

But... there are more vowels.

Absolutely. A rough representation of the IPA chart would be this XY pad morphing with the four corner vowels spanning the possible space:

VowelMorphXY.png (5.58 KiB) Viewed 1186 times

I have wondered a couple of times which filters would be best to do this with, it seems linear phase isnt quite it.

An allpole filter (the one used here) is also the natural choice in LPC, refer to this thread.

[RunBeerRun2]

Pretty cool! I like the 2-way one best, cpu-wise.

[tester]

This one is cool, thanks

Now some pitch changer and we have singing whispers and vocals

The pitch changer must be designed to understand formants of human voice as far I remember. Or the formnat scheme could include pitch changing via shifting the filter values (correlated with input frequency change).

[cyto]

Very cool examples, Martin, as always. I wonder, though, if it is really necessary to have all the coefficients declared and interpolated, etc in code. Couldn't it be just as effectively be done in green? Also, I'm assuming this isn't compatible across sample rates. What rate is it tuned to? 44.1kHz?

-cyto

[MegaHurtz]

It does sound lovely, even on my standard laptop speakers
Works nice on an impulse as well.

[infuzion]

But... there are more vowels.

Absolutely. A rough representation of the IPA chart would be this XY pad morphing with the four corner vowels spanning the possible space:

Nice 10 vowel filter! I do think you can make the XY value setting in one Code block with lots of reduction, rather than 5 modules that you have now. Tip: I'd try a Float Array to select between banks of vowel values, since you don't expect that to change too much. Get Troggie to do some SSE array compressing if you can get him to stroke the keyboard instead of his new woman long enough.

Interesting that it uses 10 poles. Usually I've seen 3, though I hope for 4 (SSE) for 10+ vowels. Google is not helping though.

The OSM is giving me a flashback:

[martinvicanek]

Thanks for your good feedback guys! Yes, the 4 way interpolator looks a little bulky with 46 inputs, hehe. It could be done in a cascade of two blocks half the size each, but that would just be cosmetics without CPU gain. Yes, it could also be done in green, and in fact that was my first thought. I discarded that idea, though, because in green, operations are not done synchronously, resulting in an inconsistent parameter set for a short period of time when you change the xy too fast. I was afraid that this transient could make the filter unstable. However, thinking again, the two pole filters are so docile it should not be a problem. Hmmm.
@Cyto: well spotted, it' for 44.1 kHz only. I'll upload a sample rate independent version later. Thanks!
@infuzion: the original filter is from musicdsp, not mine. The 10 poles are in fact 5 formants (pairs of complex conjugates). Which is still one too many for SSE Each formant is characterized by its frequency and strength (the latter is less well documented in the literature).

[infuzion]

I was thinking the data would be in green, but the XY crossover would be in Code/ASM.

[trogluddite]

Sounds great!
I'd be tempted not to bother with using all the SSE channels - the filter code is very economical, and could be really beautiful used on a poly synth - a nice multi-stage envelope for each of the two crossfade axes would make some great evolving pad sounds!
One trick does occur to me - getting the correct coefficient values to the 'corner' variables of the crossfader could be done using a routine that is executed only upon changing the selections (or from a note trigger when used in a synth).

[martinvicanek]

Okay, here is my next (and probably last) update:
- filter coefficients are handled completely in green
- coefficients are adjusted so the filter works independent of sample rate
- schematic is now better organized
Not that it'd improve on CPU or sound any different though.
I am sure there is still potential to cut down a few cycles in the filter processing. Packing 4 of the 5 filters into Mono4 has not proved efficient, however. Obviously that optimizing part is beyond me.

[martinvicanek]

a nice multi-stage envelope for each of the two crossfade axes would make some great evolving pad sounds!
One trick does occur to me - getting the correct coefficient values to the 'corner' variables of the crossfader could be done using a routine that is executed only upon changing the selections (or from a note trigger when used in a synth).

Yeah, I can almost hear it! Trog, you are so creative. I wish I had read this before, and also infuzion's comment. Looks like one should do the corners in green and the xy-mixing in code - that would leave all options open. Hmmm.