Menu

pocketsphinx 0.7- word times may be incorrect

Help
2011-04-27
2012-09-22
  • NGUYEN dang-khoa

    Dear sir,
    i use pocketsphinx_continuous version 0.7 to decode an wave file with word
    times enable. but the result i got seem to be not associate with wav file. The
    wav file has a long silence at start of file.
    Pls help me to check it !
    Thanks in advance,

     
  • Nickolay V. Shmyrev

    Sorry, I don't quite understand your question. Can you elaborate please?

     
  • NGUYEN dang-khoa

    i use this config to recognize a wave file
    pocketsphinx_continuous \
    -hmm asr/19_largevocabulary_vi.cd_ptm_5000_delinterp \
    -lm asr/models/smarthouse.lm \
    -dict asr/models/smarthouse_ascii.dic \
    -fdict asr/models/noisedict \
    -infile ${f} \
    -bestpath yes \
    -rawlogdir ${rawLogdir} \
    -time yes \
    -backtrace no 1> ${resultLogf} 2> /tmp/$$.log

    i enable switch -time to get an alignement in time domain. but when i check
    the result it seem not to be correct. the words sequence is associatae with
    transcription but not associate with signal in time domain. it mean in wav
    file the word A is from 1.5s to 2.1s but in result i get word A is from 1s to
    1.4s ....

     
  • Nickolay V. Shmyrev

    Hello

    Please provide me the data to reproduce the error. I don't see it here with
    the English model. I see that times are correct on my tests.

    Maybe your acoustic model causes this. It might be the issue with your
    training database which has silences around utterances and because of that
    phonetic segmentation using HMM is not always correct.

     
  • NGUYEN dang-khoa

    Thank you for your answer,
    here is my model and test wav file
    http://dl.dropbox.com/u/5137777/download/timealign.tar.gz
    after extract it, you run runme.sh shellscript file and then you will get
    label file which be extracted from log-result by my script.
    that is label format of Audacity,
    i used audacity to open wave file and imported the label file abow.
    the word sequence is almost correct but it is not associcate with audio
    segmentation
    Thanks in advance.

     
  • Nickolay V. Shmyrev

    Hello

    This bug has been just fixed in trunk. Please update

    Thank you for the bug report.

     

Log in to post a comment.