It's a matter of putting pi one place or the other, so I don't see
either one making more sense than the other. I simply harped on it
because I thought it was the Universal Definition (TM), but I was
wrong (again)...
--
% Randy Yates % "She's sweet on Wagner-I think she'd die for Beethoven.
%% Fuquay-Varina, NC % She love the way Puccini lays down a tune, and
%%% 919-577-9882 % Verdi's always creepin' from her room."
%%%% <yates@ieee.org> % "Rockaria", *A New World Record*, ELO
Reply by Jerry Avins●January 18, 20042004-01-18
Randy Yates wrote:
> Jerry Avins <jya@ieee.org> writes:
>
>>[...]
>>In DSP, the form with pi is more generally useful and far more
>>common.
>
>
> Jerry et al.,
>
> Bet you thought this was dead, huh?
>
> I just noticed tonight that Pawlak et al. define "sinc(t) = sin(t)/t"
> in their article "Postfiltering Versus Prefiltering for Signal
> Recovery From Noisy Samples" in the December 2003 issue of IEEE
> Transactions on Information Theory. Maybe this definition is
> more widely used that you or I thought?
Until I started hanging out with you DSP guys, it was the only one I
knew. In DSP, the pi scaling makes sense.
Jerry
--
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������
Reply by Randy Yates●January 18, 20042004-01-18
Jerry Avins <jya@ieee.org> writes:
> [...]
> In DSP, the form with pi is more generally useful and far more
> common.
Jerry et al.,
Bet you thought this was dead, huh?
I just noticed tonight that Pawlak et al. define "sinc(t) = sin(t)/t"
in their article "Postfiltering Versus Prefiltering for Signal
Recovery From Noisy Samples" in the December 2003 issue of IEEE
Transactions on Information Theory. Maybe this definition is
more widely used that you or I thought?
--
% Randy Yates % "Midnight, on the water...
%% Fuquay-Varina, NC % I saw... the ocean's daughter."
%%% 919-577-9882 % 'Can't Get It Out Of My Head'
%%%% <yates@ieee.org> % *El Dorado*, Electric Light Orchestra
Reply by Richard Owlett●January 16, 20042004-01-16
Rick Lyons wrote:
> On Thu, 15 Jan 2004 07:53:36 -0600, Richard Owlett
> <rowlett@atlascomm.net> wrote:
>
>
>>Rick Lyons wrote:
>>
>>>On Sun, 11 Jan 2004 15:22:59 -0600, Richard Owlett
>>><rowlett@atlascomm.net> wrote:
>>>
>>>
>>>
>>>>[SNIP]
>>>>Actually I suspect I need to know more about "windowing".
>>>>Any references besides a book coming out in March ;>
>>>
>>>
>>>Hi Richard,
>>>
>>> there's a decent two-part tutorial on windows at:
>>>
>>>Part I:
>>>
>>>http://www.e-insite.net/tmworld/index.asp?layout=issueTOC&pubdate=6/1/1998
>>>
>>>I just noticed the website now requires your E-mail
>>>address before they'll give the Part I article.
>>>
>>>Part II can be found at:
>>>
>>>http://www.reed-electronics.com/tmworld/index.asp?layout=article&articleid=CA187573&rid=0&rme=0&cfd=1
>>>
>>
>>A totally unbiased opinion ;? [ actually I liked it]
>>I'm wondering if a rectangular window might be fine for what I'm doing
>>at the moment as I was looking at logical ways to combine data in some
>>number of adjacent bins. My sample rate is intrinsically fixed at 4 to
>>10 times Nyquist.
>
>
> Hi Richard,
>
>
> ah shoot. I'm not sure what you're
> doing. A rectangular window is the
> same as not windowing at all.
>
> [-Rick-]
>
CONTEXT:
I have an interest in speech recognition at the VERY end user level.
I've been "bugged" by comments on comp.speech.users about how
"acoustically clean" the environment must be for good results.
Past fleeting casual contact with DSP suggested that should not be
such a significant problem.
That led to an interest in characteristics of speech. My investigation
has meandered into paths that may not have much correlation to
anything in speech recognition. It's an advantage of avocational
investigations -- you don't have to produce "useful results" :}
I wanted to start from a reproducible base.
I have a CD set of Alexander Scourby reading the _Bible_ .
This gives me many hours of one voice recorded under studio conditions
sampled at 44kHz.
TASK:
Examine speech to get a 'feel' of how it varies in time.
METHOD:
Plot a 3D surface of frequency spectrum as function of time.
[ I can't get feel of spectrograms that seem to be popular in the
speech recognition field, so I'm doing it my way ;]
To create these plots:
1. do FFT on chunk of data (currently 1 second chunks), save it
2. repeat on chunk offset by some constant (currently 50 msec)
3. plot
So really my question should be restated.
If I'm interested in frequencies from ~300 Hz to ~5 kHz, do I have any
need to apply a window function.
Also, my frequency resolution is excessive due to the externally
determined sample rate. I think some sort of running average over
adjacent bins might clean things up visually without destroying too
much information. Also considering plotting frequency on something
approaching a log scale.
Clearer, or just muddied waters more ?
Reply by Rick Lyons●January 15, 20042004-01-15
On Thu, 15 Jan 2004 07:53:36 -0600, Richard Owlett
<rowlett@atlascomm.net> wrote:
>Rick Lyons wrote:
>> On Sun, 11 Jan 2004 15:22:59 -0600, Richard Owlett
>> <rowlett@atlascomm.net> wrote:
>>
>>
>>>[SNIP]
>>>Actually I suspect I need to know more about "windowing".
>>>Any references besides a book coming out in March ;>
>>
>>
>> Hi Richard,
>>
>> there's a decent two-part tutorial on windows at:
>>
>> Part I:
>>
>> http://www.e-insite.net/tmworld/index.asp?layout=issueTOC&pubdate=6/1/1998
>>
>> I just noticed the website now requires your E-mail
>> address before they'll give the Part I article.
>>
>> Part II can be found at:
>>
>> http://www.reed-electronics.com/tmworld/index.asp?layout=article&articleid=CA187573&rid=0&rme=0&cfd=1
>>
>
>A totally unbiased opinion ;? [ actually I liked it]
>I'm wondering if a rectangular window might be fine for what I'm doing
>at the moment as I was looking at logical ways to combine data in some
>number of adjacent bins. My sample rate is intrinsically fixed at 4 to
>10 times Nyquist.
Hi Richard,
ah shoot. I'm not sure what you're
doing. A rectangular window is the
same as not windowing at all.
[-Rick-]
Reply by Richard Owlett●January 15, 20042004-01-15
A totally unbiased opinion ;? [ actually I liked it]
I'm wondering if a rectangular window might be fine for what I'm doing
at the moment as I was looking at logical ways to combine data in some
number of adjacent bins. My sample rate is intrinsically fixed at 4 to
10 times Nyquist.
Reply by Rick Lyons●January 15, 20042004-01-15
On Sun, 11 Jan 2004 15:22:59 -0600, Richard Owlett
<rowlett@atlascomm.net> wrote:
>What is 'sinc' function and why is it important.
>
>First response of "google" is "Sisters in Crime Internet Chapter" ;{
>
>OK, I did find a site with plot of 'sinc' which resembled a 'cosine
>with decreasing amplitude' -- The amplitude at 0 WAS 1 and decreased
>from there ;]
>
>Actually I suspect I need to know more about "windowing".
>Any references besides a book coming out in March ;>
Reply by Richard Owlett●January 14, 20042004-01-14
Clay S. Turner wrote:
> Hello Richard,
> Actually I did a lot of speech processing back during the 90s, so I can
> offer some help. Typically with characterizing speech there are many
> methods. So don't get too hung up on any one.
[SNIP]
> Curious what application do you have in mind or are you just looking to
> learn about this? If you have a particular app, we may be able to point to
> specific solutions.
>
Primarily, I'm just generally curious.
My current "bout of interest" was triggered by looking for 'command
and control' voice recognition software. [My pastor and a member of
the congregation have significant sight problems.]
I'm disappointed in what seems to be available. There seems to be
*less* available now than ~20 yrs ago [ on a relative scale ]. It
seems there is only one currently available commercial package --
Dragon Naturally Speaking in misc editions. It is MUCH too oriented to
continuous speech. Great for dictating "great american novel", but
otherwise useless IMHO ;{
A secondary frustration is that much too much processing horsepower
and a much too ideal acoustic environment seems to be required. I have
a "gut feel" that something is intrinsically wrong with the current
approach to speech recognition.
EG
Humans have no problem "recognizing" a wide variety of accents.
Lossy compression yields intelligible speech.
Having no academic or other reputation to risk, I wish to 'ask'
questions of a [shall we say] aberrant nature.
There are two areas in which I think I *MAY* have a chance of
contributing something useful:
1. Acoustic environment
Recent comments on "spectral subtraction" seem to indicate that there
is a possibility that I might perhaps may be thinking in a useful
direction [ did I include enough 'weasel words' ;?]
2. Speaker independence
This is a lower probability, but I suspect investigation of *WHY*
lossy speech compression is intelligible may prove fruitful.
POSSIBLY OT
Can anyone point me to a *WINDOWS* program/DLL that accepts a WAV file
and yields a stream of "characters" representing allophones &/or phonemes?
Reply by Adrian Hey●January 13, 20042004-01-13
Richard Owlett wrote:
> Jerry Avins wrote:
>
>> Randy Yates wrote:
>>
>>
>>> Sometimes it is good to simplify, but I don't think this is one
>>> of them.
>>
>>
>> That's why I thanked you for calling me on it, and described myself as
>> having "oversimplified".
>>
>> Jerry
>
> Mr. turner referenced http://mathworld.wolfram.com/SincFunction.html
> which shows that "authorities" have used both definitions.
>
> [ posted only to prove that I do follow links, even if I tend to fade
> in threads I've triggered ;]
If you look carefully this isn't quite what they say..
sinc(x)=sin(x)/x
(as I was always taught in my maths, physics and EE lectures)
The alternative definition given here uses a pi suffix (which
I cannot reproduce in a usenet post). Of course wolframs word
ain't necessarily the final word on this issue. Other "authorities"
(such as the author of Clays second link) might well have defined..
sinc(x)=sin(pi.x)/(pi.x)
which is a darned shame IMO, considering how much mathematicians,
scientists and engineers value precision. Now nobody knows for sure
what the heck sinc is, so it's best to avoid using it altogether.
I wish whoever started this trend chose a different name for the pi
version.
Regards
--
Adrian Hey
Reply by Clay S. Turner●January 13, 20042004-01-13
"Richard Owlett" <rowlett@atlascomm.net> wrote in message
news:10063q9gbm3a041@corp.supernews.com...
Hello Richard,
It is okay to refer to me as Clay.
Mr. Turner sounds so formal and putoffish. I'm a pretty easy going guy who
gets dirty on weekends playing with horses. I've been known to do my share
of pasture skiing.
And the multiple defn thing with the Sinc function is just the start. There
are several "standard" conventions for scaling the Fourier transform. If you
recall there are at least four standard ways to scale Maxwell's equations,
but you still end up with constants somewhere. So it is the same situation
with the constant pi in the Sinc function and the Fourier transform. In the
case of the Sinc function, the scaling just set the zeros of the function to
be at multiples of unity instead of multiples of pi. Convenient for
interpolating sampled data. And of course there is no zero at the origin.
This value is found by L'Hospital's rule.
>
> [ posted only to prove that I do follow links, even if I tend to fade
> in threads I've triggered ;]
>
I'm sure most go and look at the links even if they don't comment on them.
And I figured that since you move on to the next question that your last one
must have been answered.
Clay