Menu

Pronunciation modeling

Help
asr2010
2012-02-10
2012-09-22
  • asr2010

    asr2010 - 2012-02-10

    As far as I know pronunciation modeling in Sphinx can be ONLY done using a
    lookup table (dictionary). is there another way like writing the grapheme-to-
    phoneme rules in a function, then the trainer or the decoder call that
    function to get the corresponding phonetic transcription?
    Second question: in the dictionary, we can add more that one pronunciation
    variant for any word. So, is it possible to specify that one of the variants
    is only allowed at utterance beginning and another variant is allowed only in
    the middle of the utterance?
    3rd question: is it possible to specify the probability of each pronunciation
    variant
    instead of having an equal probability for all the variants for a given word?

     
  • Nickolay V. Shmyrev

    is there another way like writing the grapheme-to-phoneme rules in a
    function, then the trainer or the decoder call that function to get the
    corresponding phonetic transcription?

    This feature is not supported. It can be implemented quite easily, for example
    we have one implementation in the long audio aligner branch but hasn't been
    imported in trunk yet.

    Second question: in the dictionary, we can add more that one pronunciation
    variant for any word. So, is it possible to specify that one of the variants
    is only allowed at utterance beginning and another variant is allowed only in
    the middle of the utterance?

    I would separate create separate strings for such words. For example word_1
    and word_2 with different pronunciation. Language model trainer could care
    about position.

    3rd question: is it possible to specify the probability of each
    pronunciation variant

    This feature is not supported.

     
  • asr2010

    asr2010 - 2012-02-15

    Many thanks for your reply.

    I have another question. How can I specify in the dictionary that a specific
    word is not pronounced at all? In some languages, there are words that can be
    written in the text but we truncate them when reading. So theses words are
    only written but not pronounced. Shall we add these words in the dictionary
    without any phonemes?

     
  • Nickolay V. Shmyrev

    I have another question. How can I specify in the dictionary that a specific
    word is not pronounced at all? In some languages, there are words that can be
    written in the text but we truncate them when reading. So theses words are
    only written but not pronounced. Shall we add these words in the dictionary
    without any phonemes?

    You'd better remove those words from the language model training texts and
    insert them automatically in recognition result. Or you can join such words
    into next word. Word without pronunciation will create troubles for the
    recognizer.

     

Log in to post a comment.