Le 2012-02-02 à 02:36:00, Ed Kelly a écrit :
Still the problem with any window-based FFT is that we have to get enough points (e.g. 512, 1024) before we can do the analysis, so there is always a delay (44100/1024 = ~43ms) between live input and output (latency).
Heisenberg's Uncertainty Principle (in its acoustic version) is a hard theoretical limit, meaning it can't possibly be violated. All you can do is separate the analysis of frequencies so that you get frequent results for treble vs less frequent results for bass. This comes at the expense of having to customise, complicate and/or slow down FFT to make it do what you want.
Much more interesting is the sliding phase vocoder (Russell Bradford, Richard Dobson, and John ffitch, 2005) where the FFT is adjusted in each sample rather than each frame, and the latency is significantly reduced.
What's the speed of the sliding DFT ?
Note that they call it «DFT», and not «FFT», and that's presumably because it doesn't have to do much with FFT... whereas DFT is a more generic term for whatever computes harmonics on blocks of discrete signals.
The access to the sliding DFT article is reserved to IEEE members...
| Mathieu BOUCHARD ----- téléphone : +1.514.383.3801 ----- Montréal, QC