If I set the FFT size to 4096 in the help file and then set various settings then all output I get is heavily amplitude modulated, seemingly at a rate of every 2048 samples (that might be incorrect) - it seems like something is off with the windowing and/or window gain compensation - it’s possible that a scheme that would function under no transformation is failing in this circumstance, due to how the energy is spread in the frame due to the transformation, but that is just a guess.
I also hear the AM on default settings, but it’s easier to spot at higher FFT sizes, as it gets slower. At 8192 you will hear a tremolo effect.
It’s frustrating as this effect sounds promising, but it’s not possible to evaluate in its current state still.
Best results are with high overlap and large ffts. The original algorithm also had this effect, and @groma spent a certain time improving it in the latest version. I get quite good results, but always with a ring modulation vibe, but not wobble.
First question: on simple signal, I don’t get modulation at all. If you do, you might have wrong settings or wrong versions.
I have alpha 07. The results I get (which sound like ring mod for lower FFT and tremolo for larger) are consistent across the examples I’ve tried which are the ones in the help file and my own files.
I am not sure what you mean by “the original algorithm” - this is the first version I’ve received where I can load files without crashing, so I don’t have much of an aural comparison point to previous versions.
Open help file
Set FFT to 8192
Matrix bottom left
Temporal Shape src2
Turn volume up (I have to set around 400)
The tremolo effect should be totally clear, and doesn’t seem to be a consequence of the input material, as the tremolo is totally regular (not seemingly linked to activations).
I didn’t test this much before, but testing this now in alpha07, it sounds like pretty solid AM/ringmod in the default settings and all the variations for the parameters.
It seems I still get confused between AudioTransport and nmfmorph~ - the latter indeed will get some TLC and the former is the one that @groma worked on before the release. Sorry for the noise and stay tuned!
Hello, just to clarify, in this release there was work on audiotransport related to the sound, but nmfmorph was only fixed with respect to crashing, no change in sound. Audiotransport is a direct implementation of an algorithm proposed recently (hence it has an “original”). NMFMorph applies the same algorithm to NMF bases instead of audio spectra.
The modulation at large window sizes is really clear, and it is probably not due to this algorithm when there is no interpolation, so it is likely the phase synthesis, we are discussing this internally.