dnn的10个最佳WER

First, thanks to provide awesome speech recognition toolkit & this forum.

I am currently running my own 'isolated word dataset' recognition system in KALDI.

And I am successfully train this system with DNN, and get a 1-best decoding result.

Here is my script for scoring from local/score.sh首先，感谢您提供出色的语音识别工具包和本论坛。

我目前正在KALDI中运行自己的“孤立单词数据集”识别系统。

而且我已经成功地使用DNN训练了该系统，并获得了1最好的解码结果。

这是我用于从local / score.sh评分的脚本

"$cmd LMWT=$min_lmwt:$max_lmwt $dir/scoring/log/best_path.LMWT.log \ lattice-scale --inv-acoustic-scale=LMWT "ark:gunzip -c $dir/lat.*.gz|" ark:- \| \ lattice-add-penalty --word-ins-penalty=$word_ins_penalty ark:- ark:- \| \ lattice-best-path --word-symbol-table=$symtab \ ark:- ark,t:$dir/scoring/LMWT.tra || exit 1; "

I modify this script to get 10-best result by adding 'lattice-to-nbest' line like below :

"$cmd LMWT=$min_lmwt:$max_lmwt $dir/scoring/log/best_path.LMWT.log \ lattice-scale --inv-acoustic-scale=LMWT "ark:gunzip -c $dir/lat.*.gz|" ark:- \| \ lattice-add-penalty --word-ins-penalty=$word_ins_penalty ark:- ark:- \| \ lattice-to-nbest --acoustic-scale=0.1 --n=10 ark:- ark:- \| \ lattice-best-path --word-symbol-table=$symtab \ ark:- ark,t:$dir/scoring/LMWT.tra || exit 1; "

But I got nan WER from this script.

Can anyone give me clue how to get 10-best result?但是我从这个脚本中得到了WER。

谁能给我提示如何获得10个最佳结果？

 

 

if you inspect $dir/scoring/*.tra files and compare them with the previous files, you will notice, that the utterance ID is different for the
"lattice-to-nbest | lattice-best-path", for example UTT-A-1 instead of UTT-A. The program compute-wer then cannot find the reference (as it is looking for UTT-A-1) instead of UTT-A

BTW: I find the concept of nbest wer strange --haven't heard of it yet. Even if there would exist something like that, "lattice-to-nbest |
lattice-best-path" is almost certanly not doing anything better that just lattice-best-path -- lattice-to-nbest selects n best paths through the
lattice and lattice-best-path chooses the best one from these n paths, i.e. you will end up with 1-best path again (and I think it will be exactly the same as the path obtained from lattice-best-path).如果您检查$ dir / scoring / *。tra文件并将它们与以前的文件进行比较，您会注意到，
“lattice-to-nbest | lattice-best-path”的话语ID是不同的，例如UTT- A-1代替UTT-A。
然后，程序计算人员无法找到参考（因为它正在寻找UTT-A-1）而不是UTT-A

顺便说一句：我发现最好的概念很奇怪-还没有听说过。即使存在这样的情况，“lattice-to-nbest | lattice-best-path”几乎没有做得更好，只是lattice-to-nbest | lattice-best-path选择了穿过晶格的n条最佳路径然后，battery-best-path从这n条路径中选择最佳路径，即，您将再次获得1-best路径（我认为它将与从grid-best-path获得的路径完全相同）。

 

 

I think there might be a way to compute the n-best oracle WER given existing command line tools. The program lattice-oracle computes the lattice oracle WER; and if the input lattices consist of the n-best alternatives it will give you what you want. You can achieve this by piping your lattices through lattice-to-nbest | nbest-to-lattice.我认为，鉴于现有的命令行工具，也许有一种方法可以计算出n个最佳的Oracle WER 。
程序lattice-oracle计算格形oracle的WER；如果输入晶格由n个最佳选择组成，它将为您提供所需的内容 。您可以通过将晶格排列成nbest | nbest 最好到晶格。

 

 

I will try piping lattices through lattice-to-nbest | nbest-to-lattice too.Thank you ! I find n-best decoding result in *.tra files.I can manually analyze result from these files.我将尝试通过lattice-to-nbest | nbest-to-lattice管道方法处理晶格。我在* .tra文件中找到n个最佳解码结果。我可以手动分析这些文件的结果。