28 Jan
2005
28 Jan
'05
6:34 p.m.
On Fri, 2005-01-28 at 09:52, Alfons Adriaensen wrote:
On Fri, Jan 28, 2005 at 02:57:31PM +0000,
james(a)dis-dot-dat.net wrote:
If you're actually going to do this (argh) take a look at Tom's TAP
Fractal Doubler and use midpoint displacement fractal approximation to
add the randomness.
Jan
One thing I haven't been able to replace so
far is the Oberheim
OB-Tune plug-in. This was an amazingly useful plug-in that would take
an audio input and make sure it stayed in tune. It worked on guitars,
vocals, synths, whatever.
...
Is there anything like this out there at the moment for Linux?
Not that I know.
Operate in smallish chunks. Find the most
intense frequency (FFT or
such) and decide how far that is from the desired frequency. Scale
accordingly, preferably with as little distortion as possible, so pack
and crossfade sections.
You'll need two algorithms:
1. Pitch estimation
2. Granular resampling.
In fact the pitch estimation could be a simplified version of the real
thing. One common problem with pitch estimators is that they sometimes
lock to an harmonic or subharmonic of the real pitch. Suppose you allow
all notes on an equally tempered scale, then the 2nd, 3rd, 4th or 6th
harmonic will do as well as the fundamental when compared to the nearest
available note (the fifth would have an error of 0.8%, but it is rather
unlikely a pitch detector would ever pick it out).
A windowed FFT with interpolation will work except for low notes where
you would need a rather long transform, or use the phase information
from the succesive transforms in order to find the correct frequency.
Once you have the relative error, a small pitch shift can be made by
resampling small (25ms) overlapping chunks of the input. A raised cosine
window with 50 percent overlap will work fine in most cases. To avoid
some artefacts, add a little random variation to the window positions.