Menu

Calculating posterior probability?

2014-07-27
2014-07-27
  • Tejas Godambe

    Tejas Godambe - 2014-07-27

    Hi,

    I'm using sphinx3_decode for phone recognition. For each hypothesized phone, it gives segmentation, AScr and LMScr. How to calculate posterior probability using these?

    FV:si1386> WORD SFrm EFrm AScr(UnNorm) LMScore AScr+LScr AScale
    fv:si1386> <sil> 0 12 313835 -80587 233248 360804
    fv:si1386> ih 13 18 -33094 -127888 -160982 4710
    fv:si1386> n 19 24 -8296 -42074 -50370 82761
    fv:si1386> w 25 40 83734 -142252 -58518 239966
    fv:si1386> ey 41 56 478121 -130111 348010 667677
    fv:si1386> cl 57 59 -116825 -64665 -181490 -46743
    fv:si1386> sh 60 71 -8620 -158867 -167487 83337
    fv:si1386> en 72 78 -269832 -61654 -331486 -155035
    fv:si1386> ix 79 82 -163820 -104641 -268461 -114653</sil>

    Thanks.

     

    Last edit: Tejas Godambe 2014-07-27
  • rrb

    rrb - 2014-07-27

    Which posterior probability you are looking for? The overall loglikelihood of the bestpath itself is a posterior probability (except for the the normalization term in the denominator). However, if you want word or phone posterior, you need compute that on lattice. You can use SRILM lattice-tool to get the posterior probability, but for that you need to convert it to format it understands. There must be some API in sphinx to get that probability directly.

     
  • Nickolay V. Shmyrev

    There is no support for phone posteriors in CMUSphinx.

     

Log in to post a comment.