some fixes of tdnn+{xent,chain}+sMBR recipes on librispeech and add t… #960
Conversation
| ivector_period=$(cat $online_ivector_dir/ivector_period) || exit 1; | ||
| # note: subsample-feats, with negative n, will repeat each feature -n times. | ||
| ivector_opts="--online-ivectors=scp:$online_ivector_dir/ivector_online.scp --online-ivector_period=$ivector_period" | ||
| ivector_opts="--online-ivectors=scp:$online_ivector_dir/ivector_online.scp --online-ivector-period=$ivector_period" |
There was a problem hiding this comment.
What was the difference in performance due to this bug ?
There was a problem hiding this comment.
I didn't compare. I just found it accidently
There was a problem hiding this comment.
You should modify the commit message to highlight this bug, as it would be relevant to several recipes.
There was a problem hiding this comment.
Modifying the way this option is specified does not change the alignments, so I guess the option parser in c++ is immune this mis-specification.
freewym
force-pushed
the
librispeech
branch
2 times, most recently
from
ac05db4 to
79df9db
Compare
| @@ -443,6 +443,73 @@ | |||
| %WER 15.10 [ 7904 / 52343, 874 ins, 1070 del, 5960 sub ] exp/nnet3/tdnn_sp/decode_test_other_tgmed/wer_13_0.0 | |||
| %WER 16.29 [ 8528 / 52343, 828 ins, 1320 del, 6380 sub ] exp/nnet3/tdnn_sp/decode_test_other_tgsmall/wer_14_0.0 | |||
There was a problem hiding this comment.
This results file is really tough to read and compare. Would you be able to provide a brief set of important results right after you specify the command and present the detailed results after that ?
There was a problem hiding this comment.
Added a subset of the results before the full list of results
|
Please fix the spelling mistakes in your commit message. |
|
You can add a line new line in the commit message. Also please change In general you should have comments which follow this guideline. |
freewym
force-pushed
the
librispeech
branch
2 times, most recently
from
997d905 to
48bdb00
Compare
the results. Fixed a bug in steps/nnet3/align.sh when supplying online-ivector-period option to nnet3-align-compiled
2 participants
…heir results. Fix an option name bug in steps/nnet3/align.sh