danijel3 / clarinstudiokaldi Goto Github PK

I think that some of the scripts can run without SRILM tool. However, in local_clarin/clarin_pl_data_prep.sh there is a ngram command, that comes from SRILM, as far as I understand (but it does not perform any advanced task)

The corpus is said to be created for commercial and non-commercial use. Maybe it would be nice to allow the scripts to run with only open-licence tools too? The SRILM in this place is probably the only one limitation.

In the paper it is said that

The corpus consists of 317 speakers recorded in 554 sessions

and

56 random sessions were chosen to be in the test set, and the remaining 499 session were stored in the test [mistake, it is train] set.

Does it mean that the overfitting in terms of speakers was possible? I mean the same speaker can be in train and test set. The results don't seem reliable...

Is there any information about speaker ID in database? Then it could be possible to test the baseline system truly

Which version of Kaldi should currently be used with ClarinStudio setup?
Is it still the one "pulled from the github on Aug, 11. 2017"?

Hello @danijel3 ! Thanks for the great project!
I want to try to perform speech recognition for polish. How can I get transcriptions having only audio recordings of speech?
Thanks!

phonetisaurus_apply, that is used in clarin_prepare_dict.sh, was renamed into phonetisaurus-apply on 29.08.2017

AdolfVonKleist/Phonetisaurus@46bb7b7