Q: "Spectral domain"

[DarkStar]

I do not quite understand what this means. Could someone explain it?

For MSpectralDynamics, the video says that it extracts individual frequencies and processes each separately". What, up to 20,000 of them?

And, for the Multiband plug-ins we have:

Spectrum crossover
Spectrum crossover is the first of the spectral crossovers. It splits the signal into individual frequencies, analyzes their levels and sends the frequencies with highest level into the highest band etc. It marks each frequency with its level (as you can see on the dB X axis in the crossover band editor) and puts it into the proper band. The crossover is linear-phase and fully transparent.

Again, how many separate frequencies and how is the number determined?

[cron]

The simple answer is that (IIRC) the number of individual frequencies it works with is half the window size. E.g a window size of 4096 = 2048 points in the frequency analysis.

The longer answer:

Sound is decomposed into its component frequencies, their amplitudes and phases when you go into the frequency domain. It's a fundamentally different way of representing sound compared to its opposite: the time domain (i.e the usual way sound is dealt with - .wav files etc)

Due to the way the frequency domain transform works, you're using the window size to balance frequency accuracy against time accuracy. Lower window sizes give better time accuracy but worse frequency accuracy, and vice versa. Have a play with low and high sizes and you should get a feel for the artefacts. 1024 is a good place to start on drums, all the way up to 8192 on sustained pad-like sounds.

[DarkStar]

I thought that having multiple frequency bands might be the case (in which case the video and Info pages need revision). But I cannot see "window size" controls anywhere

[cron]

Not in front of my PC at the mo. But have the MMorph manual on the iPad and it's referred to as "buffer size" in that.

[DarkStar]

Thank you - I was looking in several other MeldaProduction plug-ins and could not see it there.

[dmbaer]

The simple answer is that (IIRC) the number of individual frequencies it works with is half the window size. E.g a window size of 4096 = 2048 points in the frequency analysis.

The longer answer:

Sound is decomposed into its component frequencies, their amplitudes and phases when you go into the frequency domain. It's a fundamentally different way of representing sound compared to its opposite: the time domain (i.e the usual way sound is dealt with - .wav files etc)

Due to the way the frequency domain transform works, you're using the window size to balance frequency accuracy against time accuracy. Lower window sizes give better time accuracy but worse frequency accuracy, and vice versa. Have a play with low and high sizes and you should get a feel for the artefacts. 1024 is a good place to start on drums, all the way up to 8192 on sustained pad-like sounds.

Disclaimer: I am just in the process of studying DSP technology, so this is only a partially authoritative response. Hopefully by 2018 I can nail it.

The first part is mostly correct. The error is using the term "window". Window is not the same thing as buffer. A window is something applied (in DFT computation) to eliminate the errors that occur when dealing with the edges of the buffer where there is probably a discontinuity.

There are different kinds of windows, and it's their shape that dictates whether they favor frequency or time accuracy. Two common window types are named Hamming (also a similar one called Hann and sometimes Hanning) and Blackman. One optimizes the calculation for frequency and other for time. I'd have to look up which is which to be honest.

If I'm not mistaken, in Melda spectral separation, we have no control over window type or size and don't even know which is employed. The buffer size we do have control over, but that's not the window size.

[MeldaProduction]

Actually there's no such thing as error. Technically if you take white noise, the number of frequencies in it is actually infinite, so that's not really useful . So if you have certain signal like that, it is naturally simulated using other sine waves. That makes it sort of impossible to say "there's this frequency and that moment", it only tells you the levels and phases of all available frequencies (equal multiplies from 0 to sampling rate / 2) and then "deal with it" . Even worse is that even if you have just a sine wave, unless this sine wave is EXACTLY one of the available frequencies, it will be simulated using many other frequencies, which is where the "window" starts to be relevant. Anyways it's a loooong and ugly topic