Librispeech: Adding lattice rescoring

[omprakashsonie]

Hi Ravanelli, Following steps are mentioned on GitHub: You can improve the performance by adding lattice rescoring in this way (run it from the kaldi_decoding_script folder of Pytorch-Kaldi):

data_dir=/data/milatmp1/ravanelm/librispeech/s5/data/ dec_dir=/u/ravanelm/pytorch-Kaldi-new/exp/libri_fmllr/decode_test_clean_out_dnn1/ out_dir=/u/ravanelm/pytorch-kaldi-new/exp/libri_fmllr/

steps/lmrescore_const_arpa.sh $data_dir/langtest{tgsmall,fglarge} \ $data_dir/test_clean $dec_dir $out_dir/decode_test_clean_fglarge || exit 1;

But I don't see following:

1. exp/libri_fmllr What should be in this directory?

I see fuller in folder ~/kaldi/egs/librispeech/s5-960

2. There is no steps directory in kaldi_decoding_script folder. It has following: ~/pytorch-kaldi/kaldi_decoding_scripts$ ls cmd.sh conf decode_dnn.sh local parse_options.sh path.sh split_data.sh utils

steps/lmrescore_const_arpa.sh is in folder: ~/kaldi/egs/librispeech/s5-960

Should give full path for steps/lmrescore_const_arpa.sh ?

3. I had run upto 13 steps of Kaldi on 960hrs data

Does it mean the following steps I ran for Librispeech were on 960hrs of data and I should compare numbers with 960hrs experiment?

• Run the Kaldi recipe for librispeech at least until Stage 13 (included)
• Copy exp/tri4b/trans.* files into exp/tri4b/decode_tgsmall_train_clean_100/
• Compute the fmllr features by running the following script.
• compute aligmenents
• run the experiments with the following command: python run_exp.py cfg/Librispeech_baselines/libri_MLP_fmllr.cfg

Any help will be appreciated.

[omprakashsonie]

Hi Ravanelli, Any help will be appreciated.

[TParcollet]

Hi,

Sorry for the delay. Did you change the path in data_dir=/data/milatmp1/ravanelm/librispeech/s5/data/ dec_dir=/u/ravanelm/pytorch-Kaldi-new/exp/libri_fmllr/decode_test_clean_out_dnn1/ out_dir=/u/ravanelm/pytorch-kaldi-new/exp/libri_fmllr/

Accordingly to your setup?

[omprakashsonie]

I am just starting/

1. will change: data_dir=librispeech/s5/data/

2. The question is for dec_dir=/exp/libri_fmllr/decode_test_clean_out_dnn1/

there is no 'libri-fmllr' in directory 'exp' and no further sub-directory 'decode_test_clean_out_dnn1'

How are these created and what is stored in it?

3. Will change path once I get through step 2 out_dir=/u/ravanelm/pytorch-kaldi-new/exp/libri_fmllr/

[TParcollet]

I am sorry but I don't understand where you current are.

Did you do step one: Run the Kaldi recipe, and you are able to see all the correct dirs created on exp/ ? Such as mono, tri2, tri3 etc etc ?

[omprakashsonie]

I have completed following:

Now trying to follow these steps:

[TParcollet]

So I'm pretty sure that steps/lmrescore_const_arpa.sh is called from the Kaldi directory. You can basically call it from here OR specify the full steps/lmrescore_const_arpa.sh path (Kaldi one)

[omprakashsonie]

ok, will provide full path for steps/lmrescore...

As 'libri_fmllr' and 'decode_test_clean_out_dnn1' directories don't exist, should I create them and run lmrescore..?

[TParcollet]

No, they should be created by steps 3.

[omprakashsonie]

Thanks a lot TParcollet for your inputs.

I was looking for directory 'libri_fmllr' in pytorch-kaldi 'exp' directory.

Looks like name has changed to 'libri_MLP_fmllr'

After correcting directory: data_dir=/home/omprakash.s/pytorch-kaldi/exp/libri_MLP_fmllr

dec_dir=/home/omprakash.s/pytorch-kaldi/exp/libri_MLP_fmllr/decode_test_clean_out_dnn1

out_dir=/home/omprakash.s/pytorch-kaldi/exp/libri_MLP_fmllr/

/home/omprakash.s/kaldi/egs/librispeech/s5-960/steps/lmrescore_const_arpa.sh $data_dir/langtest{tgsmall,fglarge} $data_dir/test_clean $dec_dir $out_dir/decode_test_clean_fglarge || exit 1;

Getting following error: Error for following directories and files:

1. libri_MLP_fmllr/lang_test_tgsmall/words.txt: No such file or directory
2. Missing file /home/omprakash.s/pytorch-kaldi/exp/libri_MLP_fmllr/lang_test_tgsmall/G.fst
3. libri_MLP_fmllr/lang_test_fglarge/words.txt: No such file or directory
4. Missing file /home/omprakash.s/pytorch-kaldi/exp/libri_MLP_fmllr/lang_test_fglarge/G.carpa

After copying these from kaldi got error: score.sh: no such file /home/omprakash.s/pytorch-kaldi/exp/libri_MLP_fmllr/test_clean/text

[omprakashsonie]

Any help will be appreciated.

[TParcollet]

You are having trouble with your paths on the Kaldi side. I am sorry that I cannot help here, but you have to fix your paths. does exp/libri_MLP_fmllr/lang_test_tgsmall/G.fst exists ? or is it exp/libri_fmllr/lang_test_tgsmall/G.fst ? The best solution for you is to look at https://github.com/kaldi-asr/kaldi/blob/master/egs/wsj/s5/steps/lmrescore_const_arpa.sh and understand properly what this script requires as arguments. Then you can reorder your paths accordingly.

[omprakashsonie]

I think the issue is with step 3. May be either there is different step 3 script for lmrescore or some issue with current step 3.

It is not creating 'lang_test_tgsmall' or having file G.fst or other required files for which lmrescore_const_arpa.sh is throwing error.