GMM模型中gconst的计算

1.基本概念

2. GMM模型中gconst的计算方法：

 给定MMF的部分截图:

~h "iu5-sh+uo1"
<BEGINHMM>
<NUMSTATES> 5
<STATE> 2
<MEAN> 39
 -3.373741e+00 -1.257332e+00 4.675202e+00 6.806309e+00 7.544928e+00 -5.862546e+00 2.479396e-01 -4.737184e+00 -1.921298e+00 6.785182e+00 -2.052
804e+00 -2.672780e+00 4.356705e+00 -3.581783e+00 -9.149255e-01 -3.685783e-01 8.497165e-01 4.364856e+00 1.556599e+00 -1.195199e-01 -8.693370e-0
1 2.649265e+00 1.585031e-01 -1.236384e-01 1.129061e+00 4.908805e-01 5.308544e-02 -6.770212e-02 -2.238328e-01 6.134966e-03 -2.301288e-02 -2.269
897e-01 -1.302624e-01 -3.047233e-01 -2.335642e-01 1.616454e-01 -7.801531e-02 1.856628e-01 2.369698e-01
<VARIANCE> 39
 2.680367e+01 7.968282e+00 2.374712e+00 1.568291e+00 3.004968e+01 1.014637e+01 9.920903e-01 3.296310e+00 3.897452e+01 5.192703e+00 6.236350e+00 3.872120e+00 2.008448e+00 1.607313e-01 1.974471e-01 2.187788e-01 1.907137e-01 5.595875e-01 5.162623e-01 4.501003e-01 7.181723e-01 2.539734e+00 9.754726e-01 3.064557e-01 2.163345e-01 2.576658e-01 2.856116e-01 1.668705e-01 1.016294e-01 1.963663e-01 2.189147e-01 4.359058e-02 1.738515e-01 4.494070e-02 1.505114e+00 1.210272e-01 5.346702e-02 1.075674e-01 8.078306e-02
<GCONST> 5.645316e+01

可得信息：D=39 ；VARIANCE是D*D的协方差矩阵的对角矩阵；|*|的值就是VARIANCE下面39个方差的乘积。|*|为行列式的值

gconst=2log(2pi^(D/2)*|*| ^(1/2))

gconst是随着方差的更新不断更新的

eg：把VARIANCE的值放到file1中

cat file1

2.680367e+01 7.968282e+00 2.374712e+00 1.568291e+00 3.004968e+01 1.014637e+01 9.920903e-01 3.296310e+00 3.897452e+01 5.192703e+00 6.236350e+00 3.872120e+00 2.008448e+00 1.607313e-01 1.974471e-01 2.187788e-01 1.907137e-01 5.595875e-01 5.162623e-01 4.501003e-01 7.181723e-01 2.539734e+00 9.754726e-01 3.064557e-01 2.163345e-01 2.576658e-01 2.856116e-01 1.668705e-01 1.016294e-01 1.963663e-01 2.189147e-01 4.359058e-02 1.738515e-01 4.494070e-02 1.505114e+00 1.210272e-01 5.346702e-02 1.075674e-01 8.078306e-02
<GCONST> 5.645316e+01

awk '{var=1;for(i=1;i<=NF;i++){var*=$i}}END{print 2*log((2*3.14)^(39/2)*var^(1/2))}'  file1

结果是56.4334，与GCONST后的值一致，注意此处log是以e为底的

3. 用dnn模型做FA时，需要用到GMM模型，此时的GMM需要做相应的变换

  1）特征变为dnn训练的特征

  2）均值变为0，方差变为1e30

  3）gconst变为2*log(p(s))。p(s)表示状态s的频率

 注意此处log是以e为底的