Menu

Keyword spotting/ word spotting

Help
sol
2011-11-12
2012-09-22
  • sol

    sol - 2011-11-12

    I am looking for a specific key word recognizer method/code. I need something
    that will be able to detect just a few words - at most 5. It needs to be able
    to listen and process continuously and work online; it needs to give real time
    feedback to the user, which means that once the user says one of the keywords,
    it needs to have a response time of less than a second and a half. Two of our
    words are "uh" and "um", so we would need to add those to the dictionary if
    it's not already included, which hopefully will not alter the program's
    effectiveness. The keyword recognizer also needs to be speaker independent. It
    needs to be able to work with a program on our computer. Once the keyword
    spotting program recognizes that the speaker has uttered a keyword, it needs
    to be able to send a signal to our other program. Our program will then power
    a vibrating motor for about a second. Do you guys know of any software that
    could do this for us? This is for a project we're working on in one of our
    classes at Dartmouth College.

    Thanks!
    Solomon

     
  • Nickolay V. Shmyrev

    Hello Solomon

    I am looking for a specific key word recognizer method/code.

    Keyword spotting task is different from speech recognition task and requires
    an implementation of the specific algorithms. We have a very initial
    implementation for sphinx4 in audio aligner branch:

    http://cmusphinx.svn.sourceforge.net/viewvc/cmusphinx/branches/long-audio-ali
    gner/KeyWordSpotting/

    But you will need more work on it. I suggest you to check it out first and try
    to plug into your application. Then we could work on making performance
    better.

    There are some other keyword spotters in the net, for example from Brno:

    http://speech.fit.vutbr.cz/software/kwsviewer-interactive-viewer-keyword-
    spotting-output

    Two of our words are "uh" and "um", so we would need to add those to the
    dictionary if it's not already included, which hopefully will not alter the
    program's effectiveness.

    Those types of words are very hard to detect reliably. You will need to
    develop a specific algorithm to detect them with low false alarm rate. You
    might initially go without them and consider them later down the road.

    Once the keyword spotting program recognizes that the speaker has uttered a
    keyword, it needs to be able to send a signal to our other program.

    It's just a technical detail.

     
  • sol

    sol - 2011-11-17

    Thanks for your response. How would my group and I go about downloading the
    audio aligner branch of the KeywordSpotting project that you directed us to?
    Is there a direct download link in the browser? We don't have svn to check out
    the project, so is there a way to get around using it or do we need to
    download that first?

     
  • Yves Raimond

    Yves Raimond - 2011-11-25

    Hello!

    I have been playing with this branch on some BBC audio, and everything seems
    to work OK. Right now, it looks like it supports only one keyword, e.g.
    'quantum number' or 'benjamin britten'.

    I am guessing it needs an update to NoSkipGrammar to handle multiple keywords.
    I guess the grammar it would need to support multiple keywords would look
    like:

    -----> kw1_1 ...---> kw1_n ----\
    InitialNode /----> kw2_1 ... ---> kw2_n-----> FinalNode

    Before I try to write a patch for that, is there something similar already
    existing I could look at?

    Kind regards,
    Yves

     
  • Nickolay V. Shmyrev

    Hello

    That would be a nice addition. You need to modify the NoSkipGrammar and the
    public interface for it. If you could rework it to better fit into keyword
    spotting task that would be just great.

     
  • Yves Raimond

    Yves Raimond - 2011-11-28

    Hello!

    I may have missed something, but the following patch seems to make this branch
    work for multiple keyword spotting:

    http://pastebin.com/9vkLrEum

    Best,
    y

     

Log in to post a comment.