Shinji Watanabe

The second audio still looks too long. Can you try to chunk the audio in less than 15 seconds?

Thanks for your report. The result looks reasonable to me. The background noise and volume issues should be fixed by re-training a model with matched conditions.

Which config file are you using? @simpleoier, can you answer it for me?

LGTM! Once you updated the result, and upload the model, I'll merge it.

Thanks. WER seems to be broken. Can you check what happens?

> Yes, I'll look into it. Regarding the corpus, we (West Point/ARL) own it. THE LDC told us that we have the right to post it to openslr.org. Sounds good....

@johnjosephmorgan, is there any progress? We'll happy to help your PR. If you have some issues you could post it or even you can directly email me ([email protected]).

Thanks for your report. I found that the first forward is deterministic while the first backward and later computations are not deterministic in some cases, probably due to this issue.

@simpleoier, there are two PRs for the same database. Both of them passed the CI check. Just FYI. @shubhamphal, I'll give you more concrete instructions on how to merge your...

@chorongi, it seems that your change in the Makefile (`tools/Makefile`) causes the issue, especially for the numpy related. Please check it. Also, move the multimodal tool installation to the extra...