Script to run DNN on Librispeech

[Agrover112]

I have been exploring the run.sh file for librispeech example, and have run the entire run.sh as it is. 

I noticed stage 14 which contains nnet2 scripts is deprecated? May I know why nnet2 is deprecated and for training dnn is there are nnet3 script we should use instead if we want to train a DNN.  Also how is nnet2 and nnet3 different?

if [ $stage -le 14 ] && false; then

# This stage is for nnet2 training on 100 hours; we're commenting it out

# as it's deprecated.

# align train_clean_100 using the tri4b model

steps/align_fmllr.sh --nj 30 --cmd "$train_cmd" \

data/train_clean_100 data/lang exp/tri4b exp/tri4b_ali_clean_100



# This nnet2 training script is deprecated.

local/nnet2/run_5a_clean_100.sh

fi

[Jan Yenda Trmal]

generally, the nnet3 and especially chain systems were giving a significant improvement in WER.

--
Go to http://kaldi-asr.org/forums.html to find out how to join the kaldi-help group
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/80202c60-8202-4b5d-b7df-70506a5bb030n%40googlegroups.com.

[Agrover112]

So which of the nnet3 scripts are we required to run to achieve the similar effect as the above stage 14 script? (as in which order do we run them? Is there any other example where nnet3 has been run so that I can mdofiy and use those scripts?)

[Jan Yenda Trmal]

there might be pointers in the "main" run.sh or (if not), there should be scripts in local/chain/run_*  and perhaps local/nnet3/run_*

y.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/b35c8609-20a8-4eca-8189-c770ba825188n%40googlegroups.com.

[Agrover112]

Thanks, just to confirm the nnet2 ones are dnn and nnet3 are tdnn and cnn/lstm and tdnn models which can explain lower WER

[Jan Yenda Trmal]

I don't recall if nnet2 already had tdnn or not. But it was ordinary softmax DNN using cross-entropy criterion.

y.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/86af45cc-e907-465b-ba64-2d720072b4b0n%40googlegroups.com.