kaldi google 论坛 问题搬运（二）：aishell_train_lms.sh报错

    借用aishell模型和字典，更改了其中的语音和标注文件，我的语料比较少，只有6000+句，在运行run.sh过程中，训练语言模型时报错如下：

Not creating raw N-gram counts ngrams.gz and heldout_ngrams.gz since they already exist in data/local/lm/3gram-mincount (remove them if you want them regenerated) Iteration 1/6 of optimizing discounting parameters discount_ngrams: for n-gram order 1, D=0.600000, tau=0.675000 phi=2.000000 discount_ngrams: for n-gram order 2, D=0.800000, tau=0.675000 phi=2.000000 discount_ngrams: for n-gram order 3, D=0.000000, tau=0.825000 phi=2.000000 interpolate_ngrams: 2846 words in wordslist discount_ngrams: for n-gram order 1, D=0.600000, tau=0.900000 phi=2.000000 discount_ngrams: for n-gram order 2, D=0.800000, tau=0.900000 phi=2.000000 discount_ngrams: for n-gram order 3, D=0.000000, tau=1.100000 phi=2.000000 interpolate_ngrams: 2846 words in wordslist interpolate_ngrams: 2846 words in wordslist compute_perplexity: for history-state "", no total-count % is seen (perhaps you didn't put the training n-grams through interpolate_ngrams?) discount_ngrams: for n-gram order 1, D=0.600000, tau=1.215000 phi=2.000000 discount_ngrams: for n-gram order 2, D=0.800000, tau=1.215000 phi=2.000000 discount_ngrams: for n-gram order 3, D=0.000000, tau=1.485000 phi=2.000000 real    0m0.007s user    0m0.008s sys    0m0.001s compute_perplexity: for history-state "", no total-count % is seen (perhaps you didn't put the training n-grams through interpolate_ngrams?) real    0m0.006s user    0m0.008s sys    0m0.000s compute_perplexity: for history-state "", no total-count % is seen (perhaps you didn't put the training n-grams through interpolate_ngrams?) real    0m0.011s user    0m0.008s sys    0m0.000s Usage: optimize_alpha.pl alpha1 perplexity@alpha1 alpha2 perplexity@alpha2 alpha3 perplexity@alph3 at /home/ml/kaldi/tools/kaldi_lm/optimize_alpha.pl line 23.

    经过问题排查，发现aishell_train_lms.sh调用了kaldi-master/tools/kaldi-lm/train_lm.sh脚本进行语言模型的训练。脚本里提到，默认语料文本至少10k句，其中heldout_sent参数默认10000，猜测改小heldout_sent参数值会解决该错误，实际上无论我把该值改成多小，都是报相同的错误。在google论坛上搜了相关问题。

    dan回答：对于语料较少的情况heldout_sent参数值设定为语料语句总数的1/10，这里修改aishell_train_lms.sh中68行的heldout_sent值没有作用，因为这是在exit 0(63行)之后，后面脚本不再执行。需要直接修改tools/kaldi_lm/train_lm.sh中的heldout_sent的值，这里要不断尝试不同的值，特别需要注意的是，修改该值之后需要删除data目录下所有生成的文件再运行run.sh方能成功（我前面之所以失败应该就是这个原因），否则始终可能报相同的错误。另外dan也建议不要使用kaldi自带工具进行语言模型的训练，而是安装SRILM，利用SRILM进行语言模型的训练，效果更好。