Menu

AM training failed @ module in training CI

Help
Jake
2011-07-13
2012-09-22
  • Jake

    Jake - 2011-07-13

    I'm training AM using 120 utterances and the length of each is about 2
    minutes. There is about 60 utterances failed in training the CI models, and
    the training process hangs at the module. I check the log files under
    logdir/20.ci_hmm/, they are all in the same error "ERROR: "backward.c", line
    431: final state not reached".

    It shouldn't be senones problem because it's failed in training CI. I searched
    online, and found it's possible caused by wrong transcript. But the
    transcripts are all correct/match after I checked them. Could anyone tell
    what's the possible reasons? How could I get the AM training continue? Really
    appreciate.

     
  • Nickolay V. Shmyrev

    Maybe your files have zero energy regions and because of that feature
    extraction fails. There could be other issues caused by corrupted audio, noise
    and so on.

    Which SphinxTrain version are you using. How do you extract features.

     
  • Nickolay V. Shmyrev

    Share the model folder with me, I will take a look.

     
  • Jake

    Jake - 2011-07-18

    Hi Nickolay, I was using the version built in March 2011 when I got the
    problem. After I compiled the latest version from the trunk, the problem is
    the same. I followed Wiki tutorial to extract features as following,

    perl scripts_pl/make_feats.pl -cfg etc/sphinx_train.cfg -ctl etc/train.fileids
    

    The the zipped model file/folders (model_architecture and model_parameters) is
    https://skydrive.live.com/?cid=e2cefd2edb0c7acd&sc=documents#!/?cid=e2cefd2ed
    b0c7acd&sc=documents&uc=1&id=E2CEFD2EDB0C7ACD!119

    Please let me know if you can open it. Thanks for your time.

     
  • Nickolay V. Shmyrev

    Hi

    I was talking about whole training folder, it's not enough to see just models.

     

Log in to post a comment.