kaldi生成FST慢

 

我也遇到生成慢的问题，7000多个音素，64G内存机器，生成了10小时没变，等的心焦，也没办法。数据检测没有致命问题。

We are trying to generate fsts from single utterances. During the context creation step, the execution time for"fstmakecontextfst" is quite high (35-40 sec). Is there a way to reduce this duration?我们正在尝试根据单个话语生成fst。在上下文创建步骤中，“ fstmakecontextfst”的执行时间非常长（35-40秒）。有没有办法减少这个持续时间？

The program fstmakecontextfst was only provided for debugging and comparison reasons, if I recall correctly; it should not be used. If you see the actual graph creation code, it calls a function that composes with the context-fst, but the context-fst is dynamically generated, only the needed parts are created. Dumping it out to disk is a bad idea because it's big.如果我没有记错的话，仅出于调试和比较的原因而提供了fstmakecontextfst程序。它不应该被使用。如果看到实际的图形创建代码，它将调用一个由context-fst组成的函数，但是context-fst是动态生成的，仅创建所需的部分。将其转储到磁盘上不是一个好主意，因为它很大。

Also - if you need to compose with the context FST, there may be a command-line program "fstcomposecontext" which you can use instead, which will avoid creating the whole thing.另外-如果需要与上下文FST组合，则可以使用命令行程序“ fstcomposecontext”来代替，这样可以避免创建整个程序。

I have used fstcomposecontext and created a CLG file. However, in order to create H file I need to use "make-h-transducer" and this executable takes a label information file as input argument which fstmakecontextfst.exe produces. Is there another way to produce H-transducer?我使用了fstcomposecontext并创建了一个CLG文件。但是，为了创建H文件，我需要使用“ make-h-transducer”，并且该可执行文件将标签信息文件作为fstmakecontextfst.exe生成的输入参数。还有生成H变换器的另一种方法吗？
I am also getting this warning after calling fstcomposecontext调用fstcomposecontext后，我也收到此警告
"WARNING (fstcomposecontext:ContextFstImpl():fstext/context-fst-inl.h:107) Context FST created but there are no phone symbols: probably input FST was empty."已创建上下文FST，但没有音素符号：输入FST可能为空。”

fstcomposecontext also outputs the ilabels-info file.The recipe is described here http://kaldi.sourceforge.net/graph_recipe_test.html
and here http://kaldi.sourceforge.net/graph_recipe_train.htmlfstcomposecontext也会输出ilabels-info文件。此方案的描述在此处http://kaldi.sourceforge.net/graph_recipe_test.html
Regarding that warning - as it says, probably the FST that you input into fstcomposecontext was empty.“警告（fstcomposecontext：ContextFstImpl（）：fstext / context-fst-inl.h：107）关于该警告-正如它所说，您输入到fstcomposecontext中的FST可能为空。

Yes, we have followed the recipe given in the links. Normally we had succefsully generated HCLG.fst by using "fstmakecontextfst" and obtained phoneme boundaries by recognition. Currently, We are using the same LG.fst file and calling the command below and still getting the warning:
"fstcomposecontext --context-size=3 --central-position=1 --read-disambig-syms=disambig_phones.list \ --write-disambig-syms=disambig_ilabels.list ilabels \ $outputFolder/LG.fst $outputFolder/CLG2.fst"
Ok, it is just solved now, it was about the worng preparation of diambiguation symbol list, currently we are able to generate HCLG.fst with fstcomposecontext and doing lattice generation for alignment.在这里http://kaldi.sourceforge.net/graph_recipe_train.html
Oh - I see now that those instructions are not optimal.The source for them is in the doc/ folder. If you have time, I'd appreciate it if you could fix the documentation, and send a patch to me
with your suggested changes (svn diff > patch.txt) to me at dpovey@gmail.com. After checking it, I can add you to the project so you can commit yourself.
是的，我们已按照链接中给出的方案进行操作。通常，我们已经使用“ fstmakecontextfst”成功生成了HCLG.fst，并通过识别获得了音素边界。当前，我们使用相同的LG.fst文件并调用下面的命令，仍然收到警告：
“ fstcomposecontext --context-size = 3 --central-position = 1 --read-disambig-syms = disambig_phones.list \\ --write-disambig-syms = disambig_ilabels.list ilabels \\ $ outputFolder / LG.fst $ outputFolder / CLG2.fst”
好的，现在就解决了，它是关于消歧符号列表的错误准备，目前，我们能够使用fstcomposecontext生成HCLG.fst并进行晶格生成以进行对齐。
哦-我现在发现这些说明不是最佳的，它们的来源在doc/文件夹中。如果您有时间，请修复文档并将补丁发送给我，我们将不胜感激。
并将您建议的更改（svn diff> patch.txt）发送给我dpovey@gmail.com。检查后，我可以将您添加到项目中，以便您可以提交自己的内容。