Question about asr2.sh and its options to reproduce the librispeech_100 recipe.

[YoshikiMas]

Describe the bug
I have faced some inconveniences when trying to reproduce the asr2 recipe for librispeech_100. I would like to clarify whether my approach is appropriate or not, especially for the second one. After finding the best way, I'll be happy to make a PR.

Stage 5
In these lines, there are ${_suf}${dset}/${train_set} and ${_suf}${dset}/${train_set}_sp. Since dset comes from a previous for loop, I expect it should be removed.

Stage 6 and Stage 7
lm_train_text is data/${train_set}/text.${tgt_case}.${tgt_lang} in the default run.sh, which does not work in Stage 6. This is because ${train_set}/text.${tgt_case}.${tgt_lang} is now included in only dump/extracted if you use speed perturbation. I can set lm_train_text to dump/extracted/wavlm_large/layer21/${train_set}/text.${tgt_case}.${tgt_lang}, but do you have any simpler ideas?
A similar problem happens in the stage 7 with src_bpe_train_text and tgt_bpe_train_text.

[sw005320]

Stage 5 In these lines, there are ${_suf}${dset}/${train_set} and ${_suf}${dset}/${train_set}_sp. Since dset comes from a previous for loop, I expect it should be removed.

I think this sounds good to me.

Stage 6 and Stage 7 lm_train_text is data/${train_set}/text.${tgt_case}.${tgt_lang} in the default run.sh, which does not work in Stage 6. This is because ${train_set}/text.${tgt_case}.${tgt_lang} is now included in only dump/extracted if you use speed perturbation. I can set lm_train_text to dump/extracted/wavlm_large/layer21/${train_set}/text.${tgt_case}.${tgt_lang}, but do you have any simpler ideas? A similar problem happens in the stage 7 with src_bpe_train_text and tgt_bpe_train_text.

I think feeding dump/extracted/wavlm_large/layer21/${train_set}/text.${tgt_case}.${tgt_lang} sounds good to me.
One option would be to get the configurations (e.g., wavlm_large and layer21) from the config file, but it is tricky.

@simpleoier, do you have any idea?

[simpleoier]

Hi @YoshikiMas, thanks for noticing the problem.
Yeah, I agree with the suggested change in stage 5.
For the lm_train_text, we can simply use data/${train_set}/text, the same as what is used in asr1. We do not need to use the pattern of ${src_case} or ${tgt_case}.
I'll prepare a PR for this after I finish the test.

[sw005320]

For the lm_train_text, we can simply use data/${train_set}/text, the same as what is used in asr1. We do not need to use the pattern of ${src_case} or ${tgt_case}. I'll prepare a PR for this after I finish the test.

I see; this is an LM for the natural text, not a speech token LM.

[simpleoier]

Yes, exactly!

[YoshikiMas]

We do not need to use the pattern of ${src_case} or ${tgt_case}.

Thank you for pointing out! I confirmed that data/${train_set}/text is the same as dump/extracted/.../text.${tgt_case}.${tgt_lang}.

Maybe only the issue is src_bpe_train_text. Of course, we can directly specify it as Shinji suggested for lm_train_text.