steps/train_mono.sh

<<单音素HMM的训练流程图.vsdx>>

定义拓扑结构、参数初始化

$ gmm-init-mono --shared-phOnes=$lang/phones/sets.int "--train-feats=$feats subset-feats --n=10 ark:- ark:-|" $lang/topo $feat_dim $dir/0.mdl $dir/tree

$ wc -w data/lang_test/phones/sets.int

385 data/lang_test/phones/sets.int

$ gmm-info exp/mono0c/0.mdl

number of phones 385

number of pdfs 290

number of transition-ids 2370

number of transition-states 1165

feature dimension 39

number of gaussians 290

对高斯参数进行扰动和幅值

totgauss=1000

# 290

numgauss=`gmm-info --print-args=false $dir/0.mdl | grep gaussians | awk ‘{print $NF}‘`

# 23，每次迭代增加23个高斯

incgauss=$[($totgauss-$numgauss)/$max_iter_inc]

# 最后一次迭代为290+23*30=980

numgauss=$[$numgauss+$incgauss]

$ gmm-est --min-gaussian-occupancy=3 --mix-up=$numgauss --power=$power $dir/0.mdl "gmm-sum-accs - $dir/0.*.acc|" $dir/1.mdl 2> $dir/log/update.0.log

$ gmm-info exp/mono0a/final.mdl

number of phones 385

number of pdfs 290

number of transition-ids 2370

number of transition-states 1165

feature dimension 39

number of gaussians 982

steps/train_mono.sh