Hi Xavier, and thank you for you reply,
I will ask this question to another forum as you suggest. One
question though that another forum might not be able to address is
how specifically SMS internally deals with the analysis window size:
When I specify an analysis window of 1024 I get 1STF frames of
windowsize 1025 and fft size 513. What happens internally?
Anyhow, I actually have tried different input spectrums
- the original spectrum with phase randomization
- the original spectrum subsampled (in frequency domain) then
linearly interpolated to reconstruct it with its full number of values.
I always get the same type of audio results which I suspect might be
due to the way I choose window sizes and what I do with that 513th
sample.
Following your recommendations today I also did some trys where I
"deconvolved" the effect of the original analysis window before I
even do the linear subsampling (or the phase randomization). Still I
get the same type of artifacts which actually are not "phase
discontinuities" as I wrongly stated in my previous message. (What I
meant by that was clicks caused by discontinuities in the synthesized
audio signal ). In fact the artifacts now is that the resynthesized
signal seems to be composed of "packets" or "wavelet kernels" which
seems to indicate that the overlap add is wrong somewhere, or that my
window shapes are off or something.
I agree that a short sound can save lines of text. I have 4 audio
examples (about 184Kb each) that I can send to you directly to avoid
sending it to the whole list. (I don't have a place to post it unless
you know of one.)
1- the interpolated noise spectrum resynth with convolution by bh92
in freq domain to compensate for the 'lost' window
2- the interpolated noise spectrum resynth without convolution:
3- the original noise spectrum resynth:
originalSpec.noconv.synth_res.wav
4- the original noise spectrum resynth with convolution of bh92 in
spec domain (window applied twice but sound more or less ok as
opposed to 1 and 2)
Thank you again for your time,
Roumbaba
On 7 août 08, at 14:18, Xavier Amatriain wrote:
Hi Baba,
Sorry for the late response but I think that this discussion is
getting a bit off-topic for this mailing list as it is more a
discussion on DSP issues than on CLAM itself. I encourage you to
take the thread to the music-dsp mailing list [1] where you will
probably get much more (and quicker) feedback on general DSP
questions... Unfortunately I don't have as much time as I wished to
get to these questions that require more thinking than writing ;-)
In any case, I don't see anything fundamentally wrong in your
procedure except in the way you have decided on the input spectrum.
The idea behind applying the BH92 to the residual spectrum was
because when doing the line approximation out of few spectral
points you are "losing" the effect of the analysis window. It is
similar to what happens when you do the peak detection process in
the sinusoidal component. If you use the original spectrum you are
in fact applying the window twice, right? Or am I missing
something? As a quick test you could try doing a peak detection +
sinusoidal synthesis (without phase continuation) also on the
residual component. This should mimic the effect of what I was
proposing... more or less.
Also, what do you exactly mean when you mention phase
discontinuities? Could you post some audio examples somewhere?
Listening to the result can sometimes save a few lines of email
text :-)
X
[1] http://music.columbia.edu/mailman/listinfo/music-dsp
roumbaba wrote:
So I have *not* managed to correctly apply the bh92 window to my
modified residual spectrum and thus I have *not* eliminate phase
discontinuities at resynth time.
One thing i still do not understand is why SMS need odd analysis
window sizes and how I should handle this. I specify analysis
window size to be 1024 and internally it seems to become 1025 and
my 1STF frames are 513 in size. The fact that i do not understand
that issue might be one of the source of what I do not do right.
Here is where I am at so far. Any hint on what I do wrong or
should do otherwise is welcome of course:
- For testing purpose the only modification I do to the original
513 values of the noise spectrum is to randomize phases.
- Then I expand the 513 spectrum to a 1026 spectrum by an even
symetry across the 513.5 axis and complex conjugate of the last
513 values.
- Then I do a circular convolution of my 1026 spectrum with the
FFT of a 1026 bh92time window.
the way I compute the bh92 time window is (matlab code for now):
w1Length = 1026;
fConst=2*pi/(w1Length+1-1);
w1=[1:w1Length];
w1=.35875 -.48829*cos(fConst*w1)+.14128*cos(fConst*2*w1) -.
01168*cos(fConst*3*w1);
- When I check the real part (and the magnitude) of the ifft of
the resulting 1026 values spectrum resulting of the convolutiong,
I do see that the windowing worked and that the resulting time
signal smoothes to 0 at begining and end.
- Then I take the first 513 values of the resulting spectrum and
replace the corresponding 1STF frame in the original sdif analysis
file
Still I get phase discontinuites in the resynth signal.
What am i missing?
Thanks,
Baba
On 15 juil. 08, at 14:55, Xavier Amatriain wrote:
Hi Roumbaba, and congrats for your progress!
You are right on the source of your problem: SMSSynthesis expects
your residual to come with an analysis window and if not things
are likely to mess up.
The lines that are "guilty" for that are around SMSSynthesis.cxx:252
http://clam.iua.upf.edu/doc/CLAM-doxygen/SMSSynthesis_8cxx-
source.html#l00252
First the peaks are synthesized into a sinusoidal spectrum. Then
the two spectrums are added. Already at that point the spectrums
are supposed to have the same analysis window (BH92) and size.
The effect of that window is undone in line 261 when the global
spectral synthesis is performed.
The issue here is that you need to guarantee that both spectrum
come from a similar place before adding them... The sinusoidal
peaks are reconstructed by convolving by the transform of the
main lobe of the window (BH92) but you are reconstructing the
residual in a different way. So.... you either apply the BH92
transform to your spectrum or avoid doing that in the peak
synthesis (and then avoid multiplying by the inverse in the
global spectral synthesis). None of the two options are immediate
but I'd say the first one should be easier to work out.
Hope it helps... and if you get it to work don't forget to report
back.
roumbaba wrote:
Hello all and thanks again for your previous help,
So I have written some matlab script to perform noise spectrum
line segment approximation.
- As input the script takes an sdif file generated by analysis
with SMSConsole.
- It then reads all sdif frames, in particular the 1STF frames
containing the noise spectrums in complex form.
- It converts these complex spectrums into magPhase form
- It performs line segment approximation on the amplitudes.
To check the impact of the approximation on the quality of
resynthesis the script does the following:
- It reconstructs full noise magnitude spectrums from the line
approximations (by linear interpolation)
- It randomizes the phases
- It converts the new "smoothed" magPhase spectrums back to
complex spectrums
- It writes back the sdif file with these new "smoothed"
spectrums instead of the original raw noise spectrums.
Then I run SMSConsole to synthesize that sdif file with the
exact same parameters than for the original sdif file.
My problem is that the resulting synthesised noise sounds like
something is wrong in the synthesis overlap-add (like lots of
discontinuites in the resynthesis)
I think that this might be due to what is described in the
Serra/Smith 1990 CMJ paper concerning line segment approximation
noise resynthesis:
" ...Since the [new] phase spectrum used is not the result of an
analysis process (with windowing of a waveform, zero padding,
and FFT computation), the resulting signal does not tapper to 0
at the boundaries. This is because a phase spectrum with random
values corresponds to a phase spectrum of a rectangular-windowed
noise waveform of size N. In order to succeed in the overlap-add
resynthesis (ie, to obtain smooth transitions between frames) we
need a smoothly windowed waveform of size M, where M is the
synthesis-window length. ....
"
So what might be happening is that by default SMSConsole assumes
that the 1STF frames are *NOT* line segment approximation and
therefore does *NOT* perform that last windowing at synthesis
time. I have gone a little bit through SMS/Clam code but I
cannot find where I can change this behavior or even if that is
the default behavior. Where shoud I look in the SMS/Clam code?
Thanks,
Roumbaba
On 27 mai 08, at 23:25, Xavier Amatriain wrote:
Hi Roumbaba,
In the paper you cite it says "you can", which does not mean
"you have to" :-) Doing an approximation of the residual model
is indeed
an interesting thing to do, especially if you want to reduce
the amount of data in your transformed signal, however it is
not a must.
Note that there are many other ways to model the residual apart
from the one mentioned in that paper.
So far, in CLAM we are using the residual as is, with no
modeling or approximation. The "only" downside is that the
transformed
signal (SMS Data) is in fact larger than the original audio
when it could be much smaller with not much loss in quality. If
for
whatever reason you do need to do the residual modeling you can
look at the SpectralEnvelopeExtract processing. This processing
generates a spectral approximation (spectrum in bpf format) but
from an array of peaks, it would not be hard to modify it to work
with an input spectrum.
X
roumbaba wrote:
Hi all,
I am trying to understand how the residual spectrum gets
modeled in clam/SMS. I have read the Serra/Smith 1990 CMJ
paper and as I understand it it describes two steps:
1- substract the harmonic spectrum from the original spectrum
2- perform a line-segment approximation of the residual
spectrum obtained in 1
I have stepped through clam and SMS code and I think I can see
where step 1 gets performed:
SMSAnalysisCore::Do()
{
mSinSpectralAnalysis.Do();
mResSpectralAnalysis.Do();
...
...
...
mSynthSineSpectrum.Do();
mSpecSubstracter.Do(); /* step 1 gets performed here I think*/
}
but I cannot find where step 2 (line approximation) gets
performed. Where should I look in the code?
Thank you very much,
Cheers,
Roumbaba
ps:
Here is a quote from the paper I mentionned above:
"Approximation of the Spectral Residual
Assuming the the residual signal is quasi-stochastic, each
magnitude-spectrum residual can be approximated by its
envelope since only its shape contributes to the sound
characteristics. [...] The particular line-segment
approximation performed here is done by stepping through the
magnitude spectrum and finding local maxima in every
section, ..."
_______________________________________________
Clam-devel mailing list
[email protected]
https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/
clam-devel
_______________________________________________
Clam-devel mailing list
[email protected]
https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/
clam-devel
_______________________________________________
Clam-devel mailing list
[email protected]
https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/
clam-devel
_______________________________________________
Clam-devel mailing list
[email protected]
https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/
clam-devel
_______________________________________________
Clam-devel mailing list
[email protected]
https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/
clam-devel
_______________________________________________
Clam-devel mailing list
[email protected]
https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/clam-
devel
_______________________________________________
Clam-devel mailing list
[email protected]
https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/clam-devel
