Menu

experiment summury

Help
hiyassat
2011-05-12
2012-09-22
  • hiyassat

    hiyassat - 2011-05-12

    I have trained arabic model :
    Total Hours Training: 6.94759935897483
    Current Overall Likelihood Per Frame = 11.3213108687957
    number of words 1429
    number of speakers 10
    when runing perl scripts_pl/decode/slave.pl
    i obtained the following results
    TOTAL Words: 2202 Correct: 1888 Errors: 361
    TOTAL Percent correct = 85.74% Error = 16.39% Accuracy = 83.61%
    TOTAL Insertions: 47 Deletions: 46 Substitutions: 268
    using same words, but different speaker
    when test on one of the speakers used for training i got
    TOTAL Words: 2202 Correct: 2049 Errors: 201
    TOTAL Percent correct = 93.05% Error = 9.13% Accuracy = 90.87%
    TOTAL Insertions: 48 Deletions: 16 Substitutions: 137
    the variables are as follows
    $CFG_N_TIED_STATES = 1000;
    $CFG_STATESPERHMM = 5;
    $CFG_FINAL_NUM_DENSITIES = 8;
    $DEC_CFG_LANGUAGEWEIGHT = "7";
    $DEC_CFG_BEAMWIDTH = "1e-120";
    $DEC_CFG_WORDBEAM = "1e-80";
    your feed back is highly appriciated

     
  • Nickolay V. Shmyrev

    Pretty good accuracy, don't you think so?

     
  • hiyassat

    hiyassat - 2011-05-15

    In order to test the effect of the number of state per HMM, I conduct new
    experiment , i got the following results
    3 states /HMM Error = 15.28%
    5 States /HMM Error = 16.39%
    is the difference due to the language i am training in this case (Arabic) or
    due to something else ?

     
  • Nickolay V. Shmyrev

    is the difference due to the language i am training in this case (Arabic) or
    due to something else ?

    No, this is a common thing. Most of the times 3 states per hmm are better,
    there is no reason to have more unless you are using long units like
    syllables.

     

Log in to post a comment.