Menu

Question about languagemodel and pocketsphinx

Help
2010-04-19
2012-09-22
  • Samuel Kitono

    Samuel Kitono - 2010-04-19

    Hi I am a newbie at these speech recognition engine. The question is probably
    very basic.
    My question is about this http://cmusphinx.sourceforge.net/wiki/languagemode
    lhowtol link where
    a corpus.txt is used to build language model and dictionary for pocketsphinx.

    So from what I understand here the corpus.txt is used as a list of words that
    you want to use in pocketsphinx. That means if i want to create a 3000 words
    system of pocketsphinx i should put 3000 words list here?

    My next question would be if I have to use complex words or names such as
    probably asean name. Will the lmtool be able to build a proper dict and lm
    file?

    Thanks in advance, sorry if my question too noob.

     
  • Nickolay V. Shmyrev

    So from what I understand here the corpus.txt is used as a list of words
    that you want to use in pocketsphinx. That means if i want to create a 3000
    words system of pocketsphinx i should put 3000 words list here?

    No, corpus is not a list of words. It's a sample text from the domain you want
    to recognize. If you are going to recognize lectures on physics, textbook on
    physic could serve as a corpus.txt. If you are going to recognize mails,
    collection of emails cleaned from HTML will be such corpus.

    My next question would be if I have to use complex words or names such as
    probably asean name. Will the lmtool be able to build a proper dict and lm
    file?

    No, the dictionary will have to be checked manually.

     
  • Samuel Kitono

    Samuel Kitono - 2010-04-20

    No, corpus is not a list of words. It's a sample text from the domain you
    want to recognize. If you are going to recognize lectures on physics, textbook
    on physic could serve as a corpus.txt. If you are going to recognize mails,
    collection of emails cleaned from HTML will be such corpus.

    Ok so if i understand this correctly the corpus only helps to identify certain
    sentences. How about other sentences that are not included in the corpus but
    the words and phonemes are included in the dictionary? Will it still work?

    No, the dictionary will have to be checked manually.

    I presume the dictionary will be the cmudict.0.6d. This is the dictionary that
    has to be modified with the proper phonemes, i suppose?

    And the acoustic model should be available from the wsj1 so I do not have to
    do any training for english right?

    Sorry for asking too many questions and thanks for your last reply.

     
  • Nickolay V. Shmyrev

    Ok so if i understand this correctly the corpus only helps to identify
    certain sentences. How about other sentences that are not included in the
    corpus but the words and phonemes are included in the dictionary? Will it
    still work?

    Language model helps to identify possible words combinations but it doesn't
    restrict them. You can read about language models in speech recognition
    textbook or at least start with
    http://en.wikipedia.org/wiki/N-gram

    I presume the dictionary will be the cmudict.0.6d. This is the dictionary
    that has to be modified with the proper phonemes, i suppose?

    The dictionary is generated by lmtool. That was the dictionary I'm talking
    about.

    And the acoustic model should be available from the wsj1 so I do not have to
    do any training for english right?

    No training is required to recognize US English, you can just take wsj
    acoustic model provided.

     

Log in to post a comment.

Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.