Menu

AM error help

Help
SKR
2016-11-17
2016-11-18
  • SKR

    SKR - 2016-11-17

    Hello,
    I found this error during AM training:

    ERROR: "ngramsearch.c", : Failed to find DAG node corresponding to (slash s)

    How can I solve this..??

    Please help,,,,
    Thanks in advance.

     

    Last edit: SKR 2016-11-17
    • Arseniy Gorin

      Arseniy Gorin - 2016-11-17

      Are you sure you have a proper transcript format? According to the tutorial, the sentences should start with <s> and end with </s>

      If this does not help, you should provide the directory files you use to investigate the issue

       

      Last edit: Arseniy Gorin 2016-11-17
  • SKR

    SKR - 2016-11-17

    This is a sharable link for the database....

    https://drive.google.com/open?id=0B_74UylilDfCNllOSnRLRVc3ZDQ

    Please find it and help me...

     
    • Arseniy Gorin

      Arseniy Gorin - 2016-11-17

      Are you sure you use the last version of sphinx?

      I run your example. The training went fine and produced errors for just 20 files. I cannot get why they are not aligned properly (likely audio/transcript mismatch), but it is less than 5% so it's not a big deal

      I attach my taining folder and the resulting models. My config and yours seem quite similar so I believe your error is coming from the software version.

      Please use last version of sphinxtrain and pocketsphonx. You can also decode my models directly. I got on your data set:
      SENTENCE ERROR: 38.4% (173/450) WORD ERROR RATE: 20.4% (314/1547)

       
  • SKR

    SKR - 2016-11-18

    Okay sir,,,, Thank you....

     
  • SKR

    SKR - 2016-11-18

    I have latest version of sphinx.... But I am getting the error....

     
    • Arseniy Gorin

      Arseniy Gorin - 2016-11-18

      Did you install pocketsphinx? It should work, you training files were fine..

       

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.