Hanning Window Overlap Factor

Hello,

In the following statement,

"The audio signal is segmented into overlapping frames.  The overlapping
frames have a length of 0.37 seconds and are weighted by a Hanning window
with an overlap factor of 31/32."

What does the 31/32 overlap factor mean exactly?  Does it mean the windows
on either side of a frame overlap almost the entire frame (31/32 of the
frame on each side).  Or does it mean that the edges of each frame (1/32 of
the frame on each side) are overlapped by neighboring windows?

31/32 seems like a lot to overlap, so that's why I asked.  Also, I looked
online and I see overlap factors that are greater than 1, so I'm confused.

Thanks.

"bogfrog" <aj00mcgraw@gmail.com> wrote in
news:beWdnYPQw61erBrVnZ2dnUVZ_q_inZ2d@giganews.com: 

> Hello,
> 
> In the following statement,
> 
> "The audio signal is segmented into overlapping frames.  The
> overlapping frames have a length of 0.37 seconds and are weighted by a
> Hanning window with an overlap factor of 31/32."
> 
> What does the 31/32 overlap factor mean exactly?  Does it mean the
> windows on either side of a frame overlap almost the entire frame
> (31/32 of the frame on each side).  Or does it mean that the edges of
> each frame (1/32 of the frame on each side) are overlapped by
> neighboring windows? 
> 
> 31/32 seems like a lot to overlap, so that's why I asked.  Also, I
> looked online and I see overlap factors that are greater than 1, so
> I'm confused. 
> 
> Thanks.
> 
> 
> 

Start with what you are trying to achieve with the algorithm, not the 
details.  If you are trying for a spectrogram, 31/32 might be reasonable.

-- 
Scott
Reverse name to reply

On Jul 23, 9:27&#4294967295;am, "bogfrog" <aj00mcg...@gmail.com> wrote:
> Hello,
>
> In the following statement,
>
> "The audio signal is segmented into overlapping frames. &#4294967295;The overlapping
> frames have a length of 0.37 seconds and are weighted by a Hanning window
> with an overlap factor of 31/32."
>
> What does the 31/32 overlap factor mean exactly? &#4294967295;Does it mean the windows
> on either side of a frame overlap almost the entire frame (31/32 of the
> frame on each side). &#4294967295;Or does it mean that the edges of each frame (1/32 of
> the frame on each side) are overlapped by neighboring windows?
>
> 31/32 seems like a lot to overlap, so that's why I asked. &#4294967295;Also, I looked
> online and I see overlap factors that are greater than 1, so I'm confused.
>
> Thanks.

bfrog,

That is what it means. Lots of overlap. If you are doing processing on
overlapping signal segments that you are windowing segments that are
extracted, processed and put back together, 50% overlap for a Hanning
window contribues a gain of 1 between original and reconstructed
signal (unless something abut the processing imacts this). For the
same extract/process/reconstruct type of procedure, shifting by 1/32
of the analysis frame should give a gain of 16 (assuming you are using
an odd length symmetrical window that peaks at 1.0, and the frame size
is divisible by 32). Could the "overlap factor" that is greater than 1
actually refer to the resultant gain between input and output signal
levels for this type of application?

I have not talked about using overlapping windows for spectrograms,
Scott Seidman's response to your post is right on the mark. And at
this point neither Scott nor myself knows what you are trying to
acheive, so explain please.

BTW, I am surprised you have not been chastized for using the name
"Hanning" to describe the window you are talking about. Purists will
compulsively tell you that it is not the correct name.  Having learned
that name in school, used that name for 25+ years, having been
understood when I used it, and not caring what the purists think, I
will not criticize your use, but it is coming. :-)

Dirk

On Jul 23, 11:38&#4294967295;am, dbell <bellda2...@cox.net> wrote:
> On Jul 23, 9:27&#4294967295;am, "bogfrog" <aj00mcg...@gmail.com> wrote:
>
>
>
>
>
> > Hello,
>
> > In the following statement,
>
> > "The audio signal is segmented into overlapping frames. &#4294967295;The overlapping
> > frames have a length of 0.37 seconds and are weighted by a Hanning window
> > with an overlap factor of 31/32."
>
> > What does the 31/32 overlap factor mean exactly? &#4294967295;Does it mean the windows
> > on either side of a frame overlap almost the entire frame (31/32 of the
> > frame on each side). &#4294967295;Or does it mean that the edges of each frame (1/32 of
> > the frame on each side) are overlapped by neighboring windows?
>
> > 31/32 seems like a lot to overlap, so that's why I asked. &#4294967295;Also, I looked
> > online and I see overlap factors that are greater than 1, so I'm confused.
>
> > Thanks.
>
> bfrog,
>
> That is what it means. Lots of overlap. If you are doing processing on
> overlapping signal segments that you are windowing segments that are
> extracted, processed and put back together, 50% overlap for a Hanning
> window contribues a gain of 1 between original and reconstructed
> signal (unless something abut the processing imacts this). For the
> same extract/process/reconstruct type of procedure, shifting by 1/32
> of the analysis frame should give a gain of 16 (assuming you are using
> an odd length symmetrical window that peaks at 1.0, and the frame size
> is divisible by 32). Could the "overlap factor" that is greater than 1
> actually refer to the resultant gain between input and output signal
> levels for this type of application?
>
> I have not talked about using overlapping windows for spectrograms,
> Scott Seidman's response to your post is right on the mark. And at
> this point neither Scott nor myself knows what you are trying to
> acheive, so explain please.
>
> BTW, I am surprised you have not been chastized for using the name
> "Hanning" to describe the window you are talking about. Purists will
> compulsively tell you that it is not the correct name. &#4294967295;Having learned
> that name in school, used that name for 25+ years, having been
> understood when I used it, and not caring what the purists think, I
> will not criticize your use, but it is coming. :-)
>
> Dirk- Hide quoted text -
>
> - Show quoted text -

Sorry for my typing... 'impacts'

dbell <bellda2005@cox.net> wrote in news:3e1899e8-6b3d-4627-a18a-
98bf3b8addbf@m36g2000hse.googlegroups.com:

> "Hanning" to describe the window you are talking about. Purists will
> compulsively tell you that it is not the correct name.  Having learned
> that name in school, used that name for 25+ years, having been
> understood when I used it, and not caring what the purists think, I
> will not criticize your use, but it is coming. :-)
> 

You wonder whether the von Hann family spits at the Hamming family in 
passing.


-- 
Scott
Reverse name to reply

On Jul 23, 9:27&#4294967295;am, "bogfrog" <aj00mcg...@gmail.com> wrote:
> Hello,
>
> In the following statement,
>
> "The audio signal is segmented into overlapping frames. &#4294967295;The overlapping
> frames have a length of 0.37 seconds and are weighted by a Hanning window
> with an overlap factor of 31/32."

Thie only began to make sense to me after about the third reading.  In
spite of the clumsy wording, I'd say it means that you are applying a
Hann window to a block of data, computing an FFT on the windowed data,
doing something with the FFT results, and then discarding the oldest
1/32 of your incoming data, adding new samples to refill the data
buffer, applying the window, etc.

>bfrog,
>
>That is what it means. Lots of overlap. 

Yes, you are right.  I continued with the project, and all the numbers
worked out perfectly when I overlapped almost the entire width of the
windows (31/32).  So I'm sure it's right now.

>is divisible by 32). Could the "overlap factor" that is greater than 1
>actually refer to the resultant gain between input and output signal
>levels for this type of application?

I was referring to this explanation:

http://www.indiana.edu/~emusic/etext/synthesis/chapter4_pv.shtml

If you scroll down, you will see he has an overlap factor of 2.  I think
he just defines it differently.

>Scott Seidman's response to your post is right on the mark. And at
>this point neither Scott nor myself knows what you are trying to
>acheive, so explain please.

I'm trying to implement fingerprint extraction on mp3's.  So I'm doing the
FFT on each window, splitting the power spectrum into bands, computing the
energy of each band, then using the energy of each band to encode
"subfingerprints" for each window.  It's an undergrad project, hence my
confusion. :)

>BTW, I am surprised you have not been chastized for using the name
>"Hanning" to describe the window you are talking about. Purists will

So the correct thing to say is "Hann window"?   I had no idea.  You can
blame Oppenheim and Schafer. :P

So I guess the name has been changed to parallel Hamming?  That actually
is quite rude!  It always seemed strange to me that there were both Hamming
and Hanning windows

John Hadstate wrote:
>Thie only began to make sense to me after about the third reading.  In
>spite of the clumsy wording, I'd say it means that you are applying a
>Hann window to a block of data, computing an FFT on the windowed data,
>doing something with the FFT results, and then discarding the oldest
>1/32 of your incoming data, adding new samples to refill the data
>buffer, applying the window, etc.

Yup, that's about right.  I was trying to word the picture I saw of the
windows in my mind.  Sorry it wasn't clear.

bogfrog wrote:

   ...

> So I guess the name has been changed to parallel Hamming?  That actually
> is quite rude!  It always seemed strange to me that there were both Hamming
> and Hanning windows

Q: Do you like Kipling?
A: I don't know. I never kippled.

   ...

Jerry
-- 
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������

"bogfrog" <aj00mcgraw@gmail.com> wrote in
news:B8adnfakz9Rc6hrVnZ2dnUVZ_uydnZ2d@giganews.com: 

> I'm trying to implement fingerprint extraction on mp3's.  So I'm doing
> the FFT on each window, splitting the power spectrum into bands,
> computing the energy of each band, then using the energy of each band
> to encode "subfingerprints" for each window.  It's an undergrad
> project, hence my confusion. :)
> 

Is the project to extract fingerprints, or to use FFT's to extract 
fingerprints?  I don't know much about fingerprints, but I do know that 
FFT's wouldn't be my first try at an approach.

-- 
Scott
Reverse name to reply

On Jul 23, 12:01&#4294967295;pm, Scott Seidman <namdiestt...@mindspring.com>
wrote:
> dbell <bellda2...@cox.net> wrote in news:3e1899e8-6b3d-4627-a18a-
> 98bf3b8ad...@m36g2000hse.googlegroups.com:
>
> > "Hanning" to describe the window you are talking about. Purists will
> > compulsively tell you that it is not the correct name. &#4294967295;Having learned
> > that name in school, used that name for 25+ years, having been
> > understood when I used it, and not caring what the purists think, I
> > will not criticize your use, but it is coming. :-)

the reason why this purist objects is that there never was a Mr./Ms./
Dr. Hanning, yet there is (or was) a Dr. Hamming  (and he had a couple
of memorable quotes).

> You wonder whether the von Hann family spits at the Hamming family in
> passing.

this reminds me of when i used to be with Wave Mechanics.  we all
liked the name but the boss was concerned about brand name conflation
with Waves.  and there was a problem.  i think the anvil that broke
the camel's back was when Sweetwater put the 3 products of WM that
they were carrying on the page with all of the dozen or so Waves
products that they were selling.  with no break in the page.  they
were literally selling PurePitch as a Waves product.  that was
disgusting enough that today Wave Mechanics is called SoundToys.

r b-j