[music-dsp] Window presum synthesis

Theo Verelst theover at tiscali.nl
Mon Apr 23 14:18:13 EDT 2012

> And for people with a lot of time on their hands, we added the Sliding
> DFT version (analysis frame updated every sample) a few years ago -
> including a simple instance of frequency domain convolution for the
> windowing. A real-time version of this was presented at ICMC last year,
> running on a GPU.

Hm, I have "some" time on my hands but would rather exchange it for some 
money, but I'm interested in the graphics card acceleration (I know I 
have a library for NVidia's Cuda 4.1 which should plug right in the 
Jamin FFT I compiled with fftw3), is this implementations somewhat Open 
Source, or not? Also, did it give good speedup ?

Oh, and about the whole idea, apart from the main considerations and of 
course the studio path experiment, there is also a main point concerning 
sub-sample information in the sense of perfect reconstruction in the 
well known sense of being able to reconstruct the original signal from 
the samples. This for instance is important when you put the signal with 
(partial) FFT transform based processing lets say running at a 44.1 kHz 
sample rate on a DA converter with it's built in 
reconstruction/oversampling/anti-aliasing filter.

It is highly unlikely that once especially the higher frequencies pass 
though some sort of short term convolution/averaging filtering made by a 
FFT the "linear" properties and filter-alikeness with analog filters are 
such that the reconstruction-ability (and therewith also the 
re-sample-ability in propwer sinc-sense) us still intact. In short: 
please keep the FFt effects in at least the low range (because of those 
horrible "always the same" near-DC transform products) and the high 
range (because of those sampling issues) away from my CDs and Blurays! I 
like to play back the materials over a 192 kS/s 24 bit DA converter with 
excellent distortion figures and the conversion to 192 works great on 
neutrally digitized signals and decent DSP processing (analog-congruent 
filter in the sense of z-1 for "s" or such things are reasonable for 
not-to-high frequencies), but many materials have been messed with, 
which I deplore.

So it wouldn't be a bad idea to check with the better studio people (I 
mean the great ones, of course) to get a piece of their mind about how 
to create digital effect equipment which can bring back the *music* and 
wonderful sounds, and what are right ways to go about using the great 
new processign resources we have at our disposal!


More information about the music-dsp mailing list