Prev: TDM with PCM question
Next: From today's junk mail
From: tmtlib on 22 Jun 2010 22:23
Relax guys! First of all it was my first post at comp.dsp, i hope for some
understanding. I interested in Dmitry's method, because it uses
n-dimensional space embedding. It is my fault posting one-dimensional
example, that is actually too basic to argue. I think that we must
sometimes use another point of view, so that many problems can be solved in
different way. 3D space embedding is especially interesting, because there
are many algorithms, that can be used to analyze discrete signals embedded
in that space. Also there are many tasks, where I&Q quadrature signals are
used, so embedding can be made in more ways.

I can imagine our discrete signal as X coordinate of moving point in 1D
space through line. Then add one more dimension and let point "get out"
from line to 2D plane, when Y changes according to some Tau time offset.
Then "open" this trajectory to 3d space. At my point of view 3D space is
optimal for these reasons: not too heavy computationally, have
ready&optimized math in many code libraries and hardware of some DSPs and
usual 3d accelerators.

p.s. Can anybody give me a link to *.wav file or write formula
(sin(x*a+b)+...+...+...), that is difficult for todays pitch estimation
methods? And can be FFT fooled in this way?

Links to existing working EXE-s of ADMF and other pitch estimation
algorithms will be useful to compare with.

p.p.s. As i promised i release some new code and WAV files. I will review
your posts guys, theory and formulas. So please make new replies slower, i
need some time to react, maybe days =)
From: steveu on 22 Jun 2010 23:36
>Relax guys! First of all it was my first post at comp.dsp, i hope for
some
>understanding. I interested in Dmitry's method, because it uses
>n-dimensional space embedding. It is my fault posting one-dimensional
>example, that is actually too basic to argue. I think that we must
>sometimes use another point of view, so that many problems can be solved
in
>different way. 3D space embedding is especially interesting, because
there
>are many algorithms, that can be used to analyze discrete signals
embedded
>in that space. Also there are many tasks, where I&Q quadrature signals
are
>used, so embedding can be made in more ways.
>
>I can imagine our discrete signal as X coordinate of moving point in 1D
>space through line. Then add one more dimension and let point "get out"
>from line to 2D plane, when Y changes according to some Tau time offset.
>Then "open" this trajectory to 3d space. At my point of view 3D space is
>optimal for these reasons: not too heavy computationally, have
>ready&optimized math in many code libraries and hardware of some DSPs and
>usual 3d accelerators.
>
>p.s. Can anybody give me a link to *.wav file or write formula
>(sin(x*a+b)+...+...+...), that is difficult for todays pitch estimation
>methods? And can be FFT fooled in this way?

You can't fool an FFT. It merely produces a dry mathematical refactoring of
a signal in a different domain. Only algorithms which try to draw
inferences, like "X is the dominate fundamental pitch in this signal", can
be fooled.

>Links to existing working EXE-s of ADMF and other pitch estimation
>algorithms will be useful to compare with.
>
>p.p.s. As i promised i release some new code and WAV files. I will review
>your posts guys, theory and formulas. So please make new replies slower,
i
>need some time to react, maybe days =)

Steve

From: tmtlib on 23 Jun 2010 02:06
>>p.s. Can anybody give me a link to *.wav file or write formula
>>(sin(x*a+b)+...+...+...), that is difficult for todays pitch estimation
>>methods? And can be FFT fooled in this way?
>
>You can't fool an FFT. It merely produces a dry mathematical refactoring
of
>a signal in a different domain. Only algorithms which try to draw
>inferences, like "X is the dominate fundamental pitch in this signal",
can
>be fooled.

I think FFT can be fooled easily using ideal sine wave with rapidly growing
(or falling) frequency. Is not it? Of course 4096-points sliding window and
zero-padding may help.

FFT is ideal for spectrometers, where accumulating spectra allows reduce
measurement errors. Maybe it "produces a dry mathematical refactoring" for
such type of signals.
From: Vladimir Vassilevsky on 23 Jun 2010 12:39


tmtlib wrote:


>>>p.s. Can anybody give me a link to *.wav file or write formula
>>>(sin(x*a+b)+...+...+...), that is difficult for todays pitch
> estimation methods? And can be FFT fooled in this way?
>>
>>x =(1 + sin(t*W/3))*sin(W*t)
>>
> ok, i'll check it
>

Here is a couple of other signals for Dr. Terrez:

x = (a * t)*sin(W*t)

x = sin((W + a)*t) + sin((2*W + a)*t) + sin((3*W + a)*t) +
+...sin((N*W + a)*t)


W = 50...500 Hz
The pitch estimate to be computed with the processing delay of 20
milliseconds.


Vladimir Vassilevsky
DSP and Mixed Signal Design Consultant
http://www.abvolt.com

First  |  Prev  | 
Pages: 1 2 3
Prev: TDM with PCM question
Next: From today's junk mail