Menu

4-gram DMP model to ARPA

Help
Florian B.
2016-03-29
2016-03-29
  • Florian B.

    Florian B. - 2016-03-29

    Hi,

    I'm currently working on a speech recognition system using neural networks and I wanted to use the quadri-gram model from the LIUM for lattice rescoring. The problem is, I have a DMP model and I need an ARPA one in my case.
    I wanted to use sphinx_lm_convert to do the conversion but it gives me n-gram errors, probably because ngram_model doesn't support 3+ order?

    Thus, what do you recommend to do the conversion? I looked in IRSTLM and SRILM toolkits but couldn't find anything.

    Thank you in advance. Best Regards,

    Florian B.

     
    • Nickolay V. Shmyrev

      LM4G2DMP tool they used for conversion is available here:

      http://www-lium.univ-lemans.fr/en/content/lm4g2dmp

       
      • Nickolay V. Shmyrev

        Between, LIUM model is not very good, it is very targetted to broadcast data. I recommend to build your own from subtitles/wikipedia/books/other crawls.

         
  • Florian B.

    Florian B. - 2016-03-29

    Thank you for your quick answer Nickolay!

    I already looked in lm4g2dmp but it seems it only does ARPA to DMP conversion. I didn't see any reverse option, am I wrong? The problem is, the only 4-gram model available at LIUM is already a DMP file..

    Regarding using LIUM model, it's best suited for my project because I use broadcast data (ESTER corpus 2)!

    Best regards,

    Florian B.

     

    Last edit: Florian B. 2016-03-29
    • Nickolay V. Shmyrev

      Then you have to reverse-engineer their code.

      You can also drop a mail to Yannick, I think he could help.

       
      • Florian B.

        Florian B. - 2016-03-29

        I'll send him a mail, thanks!
        Alternatively, I could build the 4-gram ARPA model from the data using SRILM toolkit but I think it's going to take much longer than building a DMP-ARPA converter. But I may be wrong, I'll ask Yannick.

        Thank you for your answers again!

         

Log in to post a comment.