Menu

LIUM toolkit

Yifang Xu
2010-03-18
2014-07-29
  • Yifang  Xu

    Yifang Xu - 2010-03-18

    Does anyone know where to download LIUM speech recognition toolkit?

    Thank you,

     
  • Yannick Estève

    Yannick Estève - 2010-03-18

    Hi,

    Major parts of the LIUM ASR are available in the LIUM branch of the cmusphinx
    svn.
    But at this time, there is no documentation :(
    In fact, we are porting it into the official CMU Sphinx version. But we need
    time.

    The diarization system is available here:
    http://liumtools.univ-lemans.fr//index.php?option=com_content&task=blogcatego
    ry&id=32&Itemid=60

    -Yannick

     
  • Maria Eskevich

    Maria Eskevich - 2011-04-07

    Yannick, can I ask about some work with LIUM?

    As it is said on the wiki you can run --threshold option to get the best
    settings for your data, but I have a question about the input :
    java -Xmx2024m -jar ./lib/LIUM_SpkDiarization.jar --help
    --thresholds=1.5:2.5,2.5:3.5,250.0:300,0:3.0 --loadInputSegmentation
    --fInputMask="./sph/%s.sph" --sInputMask="./sph/%s.uem.seg"
    --sInput2Mask="./ref/%s.seg" --doTuning=2 --doCEClustering $show &> out.txt

    what is the format of the ref-file? and what is the uem.seg file? again what
    is the format? should I proiduce it myself or can produce by the tool?

    Thank you very much in advance,

    Maria

     
  • Nickolay V. Shmyrev

    Hello Maria

    I'm not sure if Yannick can answer your question here. You might want to
    contact them directly by mail or through cmusphinx-devel mailing list.

    As for ref and seg files they have same format.

    20041006_0800_0900_CULTURE 1 255965 4946 M U U Jacques_Derrida
    20041006_0800_0900_CULTURE 1 260912 85 M U U Nicolas_Demorand
    20041006_0800_0900_CULTURE 1 260997 8191 M U U Jacques_Derrida
    20041006_0800_0900_CULTURE 1 269188 3091 M U U Nicolas_Demorand
    20041006_0800_0900_CULTURE 1 272280 10497 M U U Jacques_Derrida
    20041006_0800_0900_CULTURE 1 282778 231 M U U Nicolas_Demorand
    20041006_0800_0900_CULTURE 1 282778 231 M U U Jacques_Derrida
    20041006_0800_0900_CULTURE 1 283009 5884 M U U Jacques_Derrida
    20041006_0800_0900_CULTURE 1 288893 124 M U U Nicolas_Demorand
    20041006_0800_0900_CULTURE 1 289018 4997 M U U Jacques_Derrida
    20041006_0800_0900_CULTURE 1 294015 4147 M U U Nicolas_Demorand
    20041006_0800_0900_CULTURE 1 298162 18753 M U U Jacques_Derrida
    20041006_0800_0900_CULTURE 1 316915 1640 M U U Nicolas_Demorand
    

    It's pretty self-explaining. First column is record name, then channel, then
    start and end times then male/female flag, two U items reserved and speaker
    name. You can create such file with simple software script from the data you
    have or manually if you have a lot of free labour.

     
  • Mihai Dogariu

    Mihai Dogariu - 2014-07-29

    Hi, a small adjustment to what Nickolay said: the next number after the start time represents the duration of that segment, not the end time. You can easily obtain the end time by adding the 2 numbers and they are expressed in tenths of miliseconds. Basically, you can divide these numbers by 100 and obtain their seconds representation.

     

Log in to post a comment.