1. 命令行格式
在AMI数据集中num-leaves=5000(即5000个不同的states),tot-gauss=80000(平均每个不同states 高斯模型数目是16个).
"Usage: steps/tandem/train_deltas.sh <num-leaves> <tot-gauss> <data1-dir> <data2-dir> <lang-dir> <alignment-dir> <exp-dir>"
2. 训练步骤总览
1)Train a monophone system (or use previously built triphone system) to get time alignments for data.
2)Build a “decision tree” foreach“monophone”.
3)For each seen triphone, accumulate sufficient statistics to train a single Gaussian per HMM state。得到的statistics存放在exp/tri1/treeacc文件。
3.Build “decision tree”
1) get questions
We cluster the phones to get questions.
A question is just a set of phones,Would normally be a phonetic category.
Here, just clusters based on acoustic similarity.
# preparing questions, roots file...cluster-phones $dir/treeacc $lang/phones/sets.int $dir/questions.int 2> $dir/log/questions.log || exit 1;
得到的questions如下:
2)Compiling the questions
Program “compile-questions” takes lists of phonemes... Transforms it into a C++ object (written to disk) that contains questions for each “key” in EventMap.
3)Setting tree roots
4)Building the decision tree
5) Build好后的decision tree存在文件tree里,tree里的信息如下:
4.treeacc文件内容