Spectral Processing Experiment

Spectral processing experiment affected in the frequency domain vs. time domain. I used AudioMulch to produce two initially identical arpeggiated baselines, each slightly varying (by say, small cut-off frequency and LFO Depth) to produce very subtle differences (I did this so I could do frequency arithmetic on the frequency and level differences between the L and R channels). These arpeggios are sent into DelayDots SpectrumWorx product (see attached screenshot). The MP3 starts with the unaltered arpeggio, transforms to a cool spectral enhancement (listen carefully for intermodulated beating), then gets a bit scratchy with more spectral glitching, then fades back to original. The background hiss is 'cuz I'm using a demo of the SpectralWorx plugin.

Click to see screenshot of Audiomulch setup

It is easy to do multiple things in the frequency domain that would be expensive if applied using convolution in the time domain. Duplicating frequency bands, creating side bands, delaying frequencies, etc... all only require a high resolution FFT -> {your frequency gunk} -> iFFT. Of course, the FFT/iFFT adds latency you wouldn't get with a discrete linear convolution in the FIR and block convolution approach. As you well know, if you FFT, depending how you FFT, you'll impose a sample latency of at least 2 x the nyquist, or, if blocking, particularly with a SIFT approach, FFTSize - FFTSize / Overlap Factor samples.

Of course, any zero-delay convolution algorithm only gives you 0 latency if your CPU usage increases as the algorithm is worked through -- which is why the CPU usage for longer-tail convolving reverbs is so high while short ones is less.

So, with the FFT approach, you impose a 1 time delay, and make it less expensive to do many operations (i.e. did you know that frequency multiplication is actually equivalent to a circular convolution?) -- or you can use CPU like it was made of sand and get zero latency with a zero delay time domain convolution.

Still, apparently you can have your cake and eat it too, ... with a caveat..., if I interpret what Bill Gardner wrote in 1995 in the Journal of the Audio Engineering Society, if you are happy with imposing a penalty of a block of N samples, in which case you can do your convolutoin in 34 log2(n) time! That's pretty fast! For a 3 second block at 44.1khz (132300 points), each output sample would need 428 multiplies. Fun Stuff!

Trans-modern music is all a matter of compromises my farties.

So, here's the zero delay, mucho CPU way of convolving in the time domain.

And here's a way that is 34log2(n)

Remember to check out Bill Gardner's Page.