Menu

Question_extra WordInsertion while decoding

Help
Shammur
2012-05-18
2012-09-22
  • Shammur

    Shammur - 2012-05-18

    Hello,

    When I am decoding my trained Acoustic model, I have noticed that in outputs
    it can recognise all the words but it aso put few of its own. Firstly I
    thought this issue can be solved with tuning LM weight but then after I tuned
    to get my best wer. The maximum error i find is insertion of extra words. So I
    assumed that if I increase/tune my word insertion penalty (wip) then this
    error can decrease a little bit.
    So I have tested varying -wip parameter.

    Can any one please suggest what can I do to overcome this extra insertion of
    words in the output.

    Thank You very much for all the help

    Example of o/p:
    find *** flight on delta to j f k from cincinnati
    find A flight on delta to j f k from cincinnati

     
  • Nickolay V. Shmyrev

    Can any one please suggest what can I do to overcome this extra insertion of
    words in the output.

    Insertion can be caused by multiple things, sometimes it's acoustic model
    issue and better acoustic model can help, sometimes it's a language model
    issue and you need a better language model. Sometimes it's graph construction
    issue and wip can help. Error analysis is a complex analysis of the whole
    training and testing process.

    You need to provide whole dataset you are using in order to get an answer on
    your question.

     

Log in to post a comment.