Loading ...
Sorry, an error occurred while loading the content.

Re: [marf-dev] Need Help with Speaker Identification Ap

Expand Messages
  • Serguei A. Mokhov
    ... @echo off ... set CLASSPATH=.;marf.jar ... set java=java -ea -verify -Xmx512m ... set debug= set graph= ... set spectrogram= if %debug% == -debug echo
    Message 1 of 1 , Apr 1, 2007
    • 0 Attachment
      Hi Vaibhav,

      Sorry I am resending this as my previous email below was rejected with the
      filename attached "testing.bat", so I renamed it as just "testing", please
      add the extension when you receive it. The rest of my reply is below.

      -s

      On Sun, 1 Apr 2007, Serguei A. Mokhov wrote:

      > Date: Sun, 1 Apr 2007 15:01:44 -0400 (EDT)
      > From: Serguei A. Mokhov <mokhov@...>
      > To: Vaibhav Rao <vaibhav_rao@...>
      > Cc: mokhov@..., marf-devel@...
      > Subject: Re: Need Help with Speaker Identification Ap
      >
      > Hi Vaibhav,
      >
      > Thanks for the message, see the reply inline:
      >
      > On Sat, 31 Mar 2007, Vaibhav Rao wrote:
      >
      > > Date: Sat, 31 Mar 2007 11:35:18 -0700
      > >
      > > Hi,
      > >
      > > I downloaded MARF and Speaker Identification app from
      > > sourceforge.net. I created a couple of .wav files and placed
      > > them into the testing and training folders respectively. I
      > > made corresponding changes into the speakers.txt file. Then
      > > i ran the .bat file into training mode. The program is
      > > identifying the user correctly but i have a few queries
      > > which i wanted to ask you abt:
      > >
      > > 1.)If add some more testing and training files then do i
      > > have to retrain the system again?
      >
      > You need you retrain the system whenever you add a *new training samples*
      > only (for existing or new speakers). The testing samples are supposed to
      > be "unseen" by the system to make it fair to compare the recognition
      > performance.
      >
      >
      > > 2.)As mentioned by you in one of your Help Forum's post the
      > > deafult.bat files takes too much time because it tries for
      > > all possible filter combinations. I tried editing the file
      > > myself but i couldn't make much out of it. Can you please
      > > help me out with it??
      >
      > Sure. The idea is that you simply need to remove the options you don't
      > want. I am thinking of pre-generating such files with fast and slow
      > algorithms, etc. (In fact I do that manually here when I need to
      > reasonably quickly to test things out here). The simplest .bat file would
      > be the one where you hardcode your parameters to the SpeakerIdentApp call.
      >
      > Anyways, I am attaching a simplest example of a reduced .bat file to say
      > the options '-endp -lpc -cheb' (you can plug any options of your choice
      > from the available ones).
      >
      >
      > > 3.)I have a samll doubt related with the o/p of the
      > > testing.bat file. In the field "Speaker's ID" it's giving
      > > just some random name but the "Expected User ID" returns the
      > > correct name. Can you please shed some light on this aspect.
      >
      > If I understand your question correctly:
      >
      > "Speaker's ID" relates to the "Speaker identified" field. Basically, this
      > is whom the system believes the speaker is as its first guess. It may not
      > be always correct due to a variety of factors. Matching between name and
      > ID is done from the speakers.txt file as well as what the classifier
      > gives.
      >
      > "Expected Speaker's ID" and "Expected Speaker" is what is the correct
      > expected answer the system ideally *should* give. Ideally "Speaker's ID"
      > and "Expected Speaker's ID" (and the corresponding names of course) should
      > match -- this is used as the recognition perfomance metric -- how many
      > correct idenitifications vs. how many incorrect the system has done. In
      > the .bat file I attached I picked the config that gives me best ratio for
      > now of 83% of recognition in its default settings. By tweaking parameters
      > to algorithms you can achieve better or worse results.
      >
      > "Second Best ID" and "Second Best Name" is an interesting metric, where
      > you try and your first guess is wrong, likely your second guess is right.
      > Like, for example, on the phone people often mistake the voice of a father
      > and his grown up son, or, likewise, between a mother and a grown
      > up daughter, or between near-same-age brothers or sisters.
      >
      >
      > BTW, these are not produced by the .bat file, rather by the Java
      > application itself.
      >
      > Hope it helps. Let me know if you have any other issues with it.
      > Thanks!
      >
      > > Regards,
      > >
      > > Vaibhav Rao
      > >
      > > (vaibhav_rao)

      --
      Serguei A. Mokhov | /~\ The ASCII
      Computer Science Department | \ / Ribbon Campaign
      Concordia University | X Against HTML
      Montreal, Quebec, Canada | / \ Email!
    Your message has been successfully submitted and would be delivered to recipients shortly.