14 Jan
2006
14 Jan
'06
9:21 p.m.
I'm about to write a DSSI/LADSPA plugin that among other things, detunes
the signal by up to 15 cents. My understanding is that detuning is
accomplished by resampling. If that's the case, what do you do
with the time difference? Do you pad/truncate to get the same number of
samples you started out with? Wouldn't that introduce undesirables?
--
Hans Fugal ; http://hans.fugal.net
There's nothing remarkable about it. All one has to do is hit the
right keys at the right time and the instrument plays itself.
-- Johann Sebastian Bach
14 Jan
14 Jan
9:42 p.m.
On Sat, 2006-01-14 at 18:21 -0700, Hans Fugal wrote:
I'm about to write a DSSI/LADSPA plugin that among
other things, detunes
the signal by up to 15 cents. My understanding is that detuning is
accomplished by resampling. If that's the case, what do you do
with the time difference? Do you pad/truncate to get the same number of
samples you started out with? Wouldn't that introduce undesirables?
Ehrmm ... perhaps I misunderstand you, but the art of resampling is all
about keeping undesireable results at a minimum.
If you stretch your input, you'll end up with more samples ...
But hey ... Could you rephrase the question?
--
10:18 p.m.
On Sun, 15 Jan 2006 at 02:42 +0100, Jens M Andreasen wrote:
On Sat, 2006-01-14 at 18:21 -0700, Hans Fugal wrote:
Happy to. I have no doubt that the resampling part will go smoothly,
just to clarify.
Unless I'm wrong (in which case I'd like to know), resampling
necessarily involves changing the duration if it's played back at the
original sample rate. In big numbers, when I resample 1 second of
44.1KHz audio to 22050Hz, I end up with 22050 samples. That means one
second of 22050Hz audio, or 1/2 second of 44.1KHz audio. This plugin
will take as input a block of n samples and return n samples, but if
downsampling reduces the number of samples, what am I to do with the
extra time? When we're only talking a few cents, there won't be a lot of
samples, but can't even a few 0s result in clicks?
You mentioned stretching, perhaps that's the solution. Is it better to
stretch to the original size after resampling, or to prestretch and then
resample?
If I've got pitch detuning all wrong, I'd like to know that too. :-)
--
Hans Fugal ; http://hans.fugal.net
There's nothing remarkable about it. All one has to do is hit the
right keys at the right time and the instrument plays itself.
-- Johann Sebastian Bach
I'm about to write a DSSI/LADSPA plugin that
among other things, detunes
the signal by up to 15 cents. My understanding is that detuning is
accomplished by resampling. If that's the case, what do you do
with the time difference? Do you pad/truncate to get the same number of
samples you started out with? Wouldn't that introduce undesirables?
Ehrmm ... perhaps I misunderstand you, but the art of resampling is all
about keeping undesireable results at a minimum.
If you stretch your input, you'll end up with more samples ...
But hey ... Could you rephrase the question? 
10:23 p.m.
[Hans Fugal]
I'm about to write a DSSI/LADSPA plugin that among
other things, detunes
the signal by up to 15 cents. My understanding is that detuning is
accomplished by resampling. If that's the case, what do you do
with the time difference? Do you pad/truncate to get the same number of
samples you started out with? Wouldn't that introduce undesirables?
It would. As far as I know, there are two common ways to do this:
You can periodically window and granulate the signal, resample the
grains and resynthesize -- that's the faster way, doing it all in the
time domain. You'll get some comb filtering from the time-stretched
phase cancelling out during grain overlap, but if you're only doing 15
cent, it might turn out just fine. There's a library out there that
does this, I can't seem to remember its name though. Anyone?
The other way I know of is to use a phase vocoder. Basically, you
periodically window and granulate and then FFT the signal, do a bit of
math on the frequency content and resynthesize. The phase vocoder
keeps track of signal phase so the comb filtering effect is no issue.
The downside is the FFT and the math required to compute phase and
amplitude; it's quite a bit heavier on the CPU. Some code that does it
this way is at quitte.de/dsp/pvoc.html but for your intended use I'd
first look at the simpler solution mentioned above.
And of course source code to these:
$ locate pitch|grep ladspa
/usr/lib/ladspa/am_pitchshift_1433.so
/usr/lib/ladspa/pitch_scale_1193.so
/usr/lib/ladspa/pitch_scale_1194.so
/usr/lib/ladspa/tap_pitch.so
should be easily obtainable. :)
Cheers, Tim
15 Jan
15 Jan
6:58 a.m.
On Sun, Jan 15, 2006 at 03:20:56AM +0100, Tim Goetze wrote:
You can periodically window and granulate the signal,
resample the
grains and resynthesize -- that's the faster way, doing it all in the
time domain. You'll get some comb filtering from the time-stretched
phase cancelling out during grain overlap, but if you're only doing 15
cent, it might turn out just fine.
Randomising the grain positions a bit gets rid of the comb filtering.
It may introduce new artefacts, but these are often 'nice'.
--
FA
9:58 a.m.
Tim Goetze <tim(a)quitte.de> writes:
[Hans Fugal]
Probably libSoundTouch?
I'm about to write a DSSI/LADSPA plugin that
among other things, detunes
the signal by up to 15 cents. My understanding is that detuning is
accomplished by resampling. If that's the case, what do you do
with the time difference? Do you pad/truncate to get the same number of
samples you started out with? Wouldn't that introduce undesirables?
It would. As far as I know, there are two common ways to do this:
You can periodically window and granulate the signal, resample the
grains and resynthesize -- that's the faster way, doing it all in the
time domain. You'll get some comb filtering from the time-stretched
phase cancelling out during grain overlap, but if you're only doing 15
cent, it might turn out just fine. There's a library out there that
does this, I can't seem to remember its name though. Anyone? The other way I know of is to use a phase vocoder.
Basically, you
periodically window and granulate and then FFT the signal, do a bit of
math on the frequency content and resynthesize. The phase vocoder
keeps track of signal phase so the comb filtering effect is no issue.
The downside is the FFT and the math required to compute phase and
amplitude; it's quite a bit heavier on the CPU. Some code that does it
this way is at quitte.de/dsp/pvoc.html but for your intended use I'd
first look at the simpler solution mentioned above.
I'd like to suggest this article:
http://www.dspdimension.com/data/html/pshiftstft.html
It is wonderfully detailed, includes all the necessary C source
code to take the basic idea and make it fly. I used this
method to write my instrument tuner by ripping out the last phase
of the algorithm (resynthesis) and only using the exact
frequency peak info. Its pretty useful, and I learned a lot
from this article. Actually, I think thats the single most useful
DSP tutorial I've ever read. It made me really grok FFT, which
is something by its own right :-)
Oh, and it is pretty simple to make this stuff work with libfftw3, just
read tuneit.c (http://delysid.org/tuneit.html).
--
CYa,
Mario
4:11 p.m.
mlang wrote:
I'd like to suggest this article:
http://www.dspdimension.com/data/html/pshiftstft.html
It is wonderfully detailed, includes all the necessary C source
code to take the basic idea and make it fly. I used this
method to write my instrument tuner by ripping out the last phase
of the algorithm (resynthesis) and only using the exact
frequency peak info. Its pretty useful, and I learned a lot
from this article. Actually, I think thats the single most useful
DSP tutorial I've ever read. It made me really grok FFT, which
is something by its own right :-)
Oh, and it is pretty simple to make this stuff work with libfftw3, just
read tuneit.c (http://delysid.org/tuneit.html).
Thanks for these excellent links, I think they will help me understand
this very well. And thanks to everyone else for their insights as well.
6688
days inactive
6689
days old
linux-audio-dev@lists.linuxaudio.org
6 comments
5 participants
participants (5)
-
fons adriaensen -
Hans Fugal -
Jens M Andreasen -
mlang -
Tim Goetze