|
|
 | | From: | Holger | | Subject: | HTK - FOM Calculation for word spotting performance | | Date: | 13 Dec 2004 08:49:31 -0800 |
|
|
 | Hi all,
I am trying to find out how the FOM in the HResults function of HTK
works.
The FOM is defined as follows: The FOM is calculated as follows where
it is assumed that the total duration of the test speech is T hours.
For each word, all of the spots are ranked in score order. The
percentage of true hits pi found before the i'th false alarm is then
calculated for i=1...N where N is the first integer N >=10T+0.5 .
The figure of merit is then defined as
FOM= 1/(10*T) * ( p_1 + ... + p_N + a*p_(N+1) )
where a= 10T-N is a factor that interpolates to 10 false alarms per
hour.
The problem I have is the following:
Example
In my recorded speech signal (60 Minutes long) I have 12 words
(groundtruth).
The Spotter finds 10 of these words, but also produces 4 False Alarms?
How do I calculate the FOM then? The problem is that for i>4 I cannot
calculate pi anymore.
I hope somebody can help.
Thanks a lot.
|
|
| | From: | James Salsman | | Subject: | Re: HTK - FOM Calculation for word spotting performance | | Date: | Tue, 14 Dec 2004 05:21:56 GMT |
|
|
| advances with great rapidity and threatens freedom at
many different points at the same time (crowding, rules and
regulations, increasing dependence of individuals on large
organizations, propaganda and other psychological techniques, genetic
engineering, invasion of privacy through surveillance devices and
computers, etc.) To hold back any ONE of the threats to freedom would
require a long different social struggle. Those who want to protect
freedom are overwhelmed by the sheer number of new attacks and the
rapidity with which they develop, hence they become pathetic and no
longer resist. To fight each of the threats separately would be
futile. Success can be hoped for only by fighting the technological
system as a whole; but that is revolution not reform.
131. Technicians (we use this term in its broad sense to describe all
those who perform a specialized task that requires training) tend to
be so involved in their work (their surrogate activity) that when a
conflict arises between their technical work and freedom, they almost
always decide in favor of their technical work. This is obvious in the
case of scientists, but it also appears elsewhere: Educators,
humanitarian groups, conservation organizations do not hesitate to use
propaganda or other psychological techniques to help them achieve
their laudable ends. Corporations and government agencies, when they
find it useful, do not hesitate to collect information about
individuals without regard to their privacy. Law enforcement agencies
are frequently inconvenienced by the constitutional rights of suspects
and often of completely innocent persons, and they do whatever they
can do legally (or sometimes illegally) to restrict or circumvent
those rights. Most of these educators, government officials and law
officers believe in freedom, privacy and constitutional rights, but
when these conflict with their work, they usually feel tha
|
|
| | From: | James Salsman | | Subject: | Re: HTK - FOM Calculation for word spotting performance | | Date: | Tue, 14 Dec 2004 03:35:05 GMT |
|
|
| Holger,
You want to ask your question on the comp.speech.research
newsgroup, where I'm crossposting it for you, or better yet
on the HTK-Users email list:
http://htk.eng.cam.ac.uk/mailing/subscribe_mail.shtml
Sincerely,
James
> Hi all,
> I am trying to find out how the FOM in the HResults function of HTK
> works.
> The FOM is defined as follows: The FOM is calculated as follows where
> it is assumed that the total duration of the test speech is T hours.
> For each word, all of the spots are ranked in score order. The
> percentage of true hits pi found before the i'th false alarm is then
> calculated for i=1...N where N is the first integer N >=10T+0.5 .
> The figure of merit is then defined as
>
>
> FOM= 1/(10*T) * ( p_1 + ... + p_N + a*p_(N+1) )
> where a= 10T-N is a factor that interpolates to 10 false alarms per
> hour.
>
> The problem I have is the following:
> Example
> In my recorded speech signal (60 Minutes long) I have 12 words
> (groundtruth).
> The Spotter finds 10 of these words, but also produces 4 False Alarms?
>
> How do I calculate the FOM then? The problem is that for i>4 I cannot
> calculate pi anymore.
> I hope somebody can help.
>
> Thanks a lot.
|
|
|
