I've put together an FFT abstraction which inserts 3 undertones, the fundamental, and 11 overtones based on incoming pitch. The user is able to superimpose harmonic series of different pitches by clicking "+ series," which may be of interest to those who work with more "klang" orientated sounds. The abstraction works with a large block size in order to work with individual frequency components. It's my first attempt at putting together something remotely musical with FFT objects, so any feedback would be greatly appreciated.

Ricky

edit: I might add that the latency that comes with using a large block size is not an issue for me.

Cool,

I look forward to checking it out tonight.

Great. You will need the [tabletool] object if you do not have it already.

http://williambrent.conflations.com/pages/research.html#tabletool

Any thoughts folks?

didn'T look into it jet.. but did you account for bucket smearing? and maybe calculate the true frequency? (i hate fft):
but well a blocksize of 1s should be quiet accurate.. but you know the blocksize is always 2^x so i guess its really 32768 samples

@slur said:

did you account for bucket smearing? and maybe calculate the true frequency?

No, I did not account for bucket smearing. I'm still new to FFT, so my patch is quite simple really (looking for direction from those in the know).

I found this link which may be of use: http://www.mstarlabs.com/apeng/integration/harmon.html in terms of explaining potential issues and possible solutions. I assume by following their proposed solution, I can reduce smearing, but like you say, a block size of 1s should be quite accurate, and also, aesthetically, smearing may produce some interesting sounds?

I am interested in calculating "true frequency" if you can elaborate, maybe suggest a method of doing so?

You'll have to excuse my mathematical ignorance, but can you elaborate on 2 to the power of x = 32768?

Thanks slur,

Ricky

Hey Ricky,

it's a very fine patch, although i've got no clue what it is doing. However a couple of remarks: it seems to have very huge cpu load (cpu was going to 60%) and things started crackling all around. Beside the crackles, it sounded like my incoming audio was "smeared" out over time. I don't know if that's the purpose of the patch but I really liked that sound.

D.

Wow. It's not eating that much CPU for me I'm on 30% at most. It's taking an incoming audio signal, tracking the pitch via [sigmund~], and inserting array index for 3 undertones, fundamental, and 11 overtones using [tabletool]. Pretty simple. I quite like the "smeared" sound too, I really like the idea of working with more "klang" orientated sounds. Although, if you or anyone else have any direction on how to build a parameter to reduce "smearing," I'd be keen to hear about. Oh, and maybe some ideas on how to reduce CPU.

im not good with fft too... i just know that i don'T know enough... currently thinking about building something phasevocoder based... the link from katja was http://www.dspdimension.com/admin/pitch-shifting-using-the-ft/ uses real frequencys... i think there was a link to a description how this is done... or at least its hidden in the source (im really put off by only thinking about porting a cpp to pd)

Thanks for the link, slur. I'll definitely look into this over the holiday.

Beginner's Guide: http://footils.org/2007/02/20/beginners-guide-fft-objects-pd/

Also helpful.

It's a cool patch. I wasn't sure what I was supposed to do for bit and I was using the graph like an EQ on the ADC.

Once I got it going it sounded a bit like reverse plate reverb at times. I tried a little vocal droning and it was able to generate nice Mongolian throat singing kind of harmonics. It would have been good if the upper harmonics were more consistent. They seemed to be drowned out by the lower harmonics after the harmonic table built up. Okay at the start but overtime it gets darker and the formant peaks seem to smear into a more general sound like a dark reverse plate reverb. Possibly there is a problem there or possibly I am imagining it or more likely misunderstanding what is happening. sorry

The CPU was bouncing around between 20 and 40 but mostly sitting around 28% on a Intel i3.

Glad you got some use out of it. The attenuation of higher partials is to be expected over time, I'd suggest reducing the amplitude of the more consonant partials i.e. f, 5ths, & 8ves. It is a very simple concept, and of course the [tabletool] object doesn't function exactly how I'd like it to. When you add additional series, it seems to shift the previously inserted series if the existing indexes are less than the index you intend to insert (see patch). So, lots more research to do before this patch is any good.

Edit: Of course [tabwrite] is more suitable (patch attached).

Thanks for the feedback thus far. I'll look into developing this patch in the new year.

I've uploaded a version with [tabwrite]. Merry Christmas.

A bold experiment with intriguing output, Ricky. The way you have implemented it, multiplying spectral information, is a frequency domain zero-phase filter. Theoretically, filters can only attenuate or amplify frequency components which are already in the signal. This is apart from artifacts which can be introduced, for example by not windowing the fft frames. If you want to add harmonics, you need to, well, add them (not multiply). As Slur has pointed out, you probably want to add true frequencies. That would mean the emulation of phase rotations and spectral smearing. In frequency domain, generating steady components which are not exactly a bin frequency is immensely complicated....

....while in time domain it is so easy. Why not use a bunch of oscillators whose frequencies are controlled by sigmund~'s pitch analysis? Admittedly, the result may not be so spectacular and alienating as your fft experiment.

Katja

Hi Katja,

Thanks for your feedback. I'll have a look at the materials on your website. I understand that this process would be a lot simpler in the time domain, but wanted to explore the possibilities with FFT. If you have any time domain examples, I'd be very grateful for any links you might have?

Best wishes,

Ricky

Attached is a patch [harmonosc~], which implements your idea of generating (sub)harmonics of an analysed pitch in time domain. It does the job quite faithfully, thanks to [sigmund~]. But, as expected, the sound is not very spectacular. Reminds of an old guitar synthesizer.

Anyway, comparison of [harmonosc~] with [fftea~] shows that theoretically correct implementations do not necessarily make for the most interesting results. Imagine you 'd spend months in frequency domain, then come up with an old guitar synthesizer.. But please keep on experimenting. Frequency domain is a place full of surprises.

Katja

http://www.pdpatchrepo.info/hurleur/harmonosc~.pd.zip

Thanks Katja. That's a good patch. I will keep experimenting, and I'll be sure to continue reading through your materials.

@katjav said:

If you want to add harmonics, you need to, well, add them (not multiply).

Hi Katja,

As far as I understand, I am adding the partials here. The only multiplication taking place is the multiplication of the fundamental to obtain the harmonic series associated with that particular frequency, and then I'm adding them to the table. Perhaps I am misunderstanding what you mean? Could you elaborate?

Best,

Ricky

In the subpatch [pd fft] you have these multiplications inbetween [rfft~] and [rifft~]. Real and imaginary coefficients of the input signal's spectrum are multiplied by the values in table 'fft', which acts as an amplitude spectrum of a filter in that case. It really is what they call (zero phase) filtering via frequency domain, an equivalent of convolution in time domain. Filters can not add frequency components which were not in the signal originally, they can only amplify or attenuate existing components, that is, the non-zero real- and imaginary coefficients in this case. Any spectrum coefficient which is zero will remain zero after multiplication....

For inserting new spectral components you could use [+~] instead of [*~] there, for a start. Then you are really creating new spectrum coefficients which translate to signal domain as new frequency components, albeit with rough method. But it would be wise to first normalise the real and imaginary coefficients from [rfft~], then insert new coefficients at a level which is in accordance with the signal's spectrum itself (check with [scope~] for example, because [snapshot~] will only show the first value in each block which happens to be DC component here). You'd need to auto-erase values in array 'fft' after each new pitch detect.

One more remark: blocksize should be set to a power of two, as Slur mentioned, otherwise Pd will reduce blocksize to a power of two without notice and your fft array length does not match spectrumsize.

Happy to see you're still busy with it. Good luck!

Katja

@katja said:

It really is what they call (zero phase) filtering via frequency domain, an equivalent of convolution in time domain.

Thanks for the clarification, Katja. With all things considered, perhaps a zero phase filter via the frequency domain will suffice for my current project.

@katja said:

For inserting new spectral components you could use [+~] instead of [*~] there, for a start.

This is very interesting, I will definitely pursue this sometime in the near future.

@katja said:

One more remark: blocksize should be set to a power of two, as Slur mentioned, otherwise Pd will reduce blocksize to a power of two without notice and your fft array length does not match spectrumsize.

Understood. Thank you for your advice.

Best,

Ricky