Menu

Confidence score

Help
Tony
2011-08-03
2012-09-22
  • Tony

    Tony - 2011-08-03

    the app always returns returns a confidence of 1 although the result doesn't
    match with the only recognizable sentence in .test. please check the java
    codes and config.xml for me. Thank you

    Tony

     
  • Nickolay V. Shmyrev

    You need to increase beams and word insertion probability (see voxforge_en
    configuration for example)

    You need to dump lattice and make sure it's diverse enough (has many decoding
    variants for each word). Unless lattice will be diverse confidence will not be
    precise.

     
  • Tony

    Tony - 2011-08-03

    Is there a link I could read about dump lattice and beams (i have no idea what
    they do and thus no idea why the value for voxforge_en was picked different
    from the other demos)

    Tony

     
  • Nickolay V. Shmyrev

    Is there a link I could read about dump lattice and beams

    You can find deep information about speech recognition algorithms in a book

    http://www.amazon.com/Spoken-Language-Processing-Algorithm-
    Development/dp/0130226165

    no idea why the value for voxforge_en was picked different from the other
    demos

    The values in demos are random. The values in voxforge_en were selected to
    provide a good balance between recognition accuracy using voxforge_en database
    and decoding speed.

     
  • Tony

    Tony - 2011-08-04

    I need emission probabilities of a certain sentence base on hmm alg of sphinx4
    (not the confidence score or wer in this case? cuz I don't care what sphinx4
    thinks the
    most likely result to be but rather I would like sphinx to give me the
    probability of
    the words that I want to recognize given the acoustic model). Where should i
    look into?

    Tony

     
  • Nickolay V. Shmyrev

    I need emission probabilities of a certain sentence base on hmm alg of
    sphinx4

    If you need an acoustic score it's available with the Token.getAcousticScore()
    method of the Token class. You can get the best token of the result using
    Result.getBestToken() method

     
  • Tony

    Tony - 2011-08-08
    1. How do I implement this method?

    2. Do I get an acoustic score for any sentences I want with this method? How do I tell with sentence to return scores for? What scores for onces that do not reach final state?

     
  • Nickolay V. Shmyrev

    How do I implement this method?

    Which method? The methods mentioned above are already implemented

    Do I get an acoustic score for any sentences I want with this method?

    Not sure what method are you talking about.

     
  • Tony

    Tony - 2011-08-08

    Sorry about not being clearer

    1. How do I use the getAcousticScore method in a Java script for example if I modify the confidence.java file, where and what changes do I need to make in the java file and .config.xml to obtain the score I want?

    2. I want to get the acoustic score for a specific sentence that I choose.What should I do?

     
  • Tony

    Tony - 2011-08-09
    1. I used a grammar model containing only the sentences that I wanted to recognize so it could only return this sentence with a score for the acoustic model. Is this approach correct?

    2. I modified the spelling of some words in this grammar and then did corresponding modification in the dictionary file so they matches but all the words i modified in the grammar showed up as missing from the dictionary although they do exist there. What could be the problem? I don't need to retrain the acoustic models right?

     
  • Nickolay V. Shmyrev

    1. How do I use the getAcousticScore method in a Java script for example if
      I modify the confidence.java file, where and what changes do I need to make in
      the java file and .config.xml to obtain the score I want?

    It's too much to explain in a single reply. Can you be more specific in your
    questions?

    1. I want to get the acoustic score for a specific sentence that I
      choose.What should I do?

    You recognize the sentence using grammar and get the score

    1. I used a grammar model containing only the sentences that I wanted to
      recognize so it could only return this sentence with a score for the acoustic
      model. Is this approach correct?

    Yes

    1. I modified the spelling of some words in this grammar and then did
      corresponding modification in the dictionary file so they matches but all the
      words i modified in the grammar showed up as missing from the dictionary
      although they do exist there. What could be the problem? I don't need to
      retrain the acoustic models right?

    Most likely you forgot to clean the old build.

     

Log in to post a comment.