dimenstion error

297 views
Skip to first unread message

Yu Beomgon

unread,
Jul 11, 2019, 1:41:06 AM7/11/19
to kaldi-help
I trained voxforge.
then I want to test with my audio using trained model.

I got a help from this blog

but dimension error happend.
trying to find, but difficult.

can anybody help me?
thanks


beomgon@beomgon-ThinkStation-P520:~/kaldi/egs/voxforge/s5$ ./test_run.sh
compute-mfcc-feats --write-utt2dur=ark,t:exp/make_mfcc/test/utt2dur.1 --verbose=2 --config=conf/mfcc.conf scp:transcriptions/test.scp ark,scp:transcriptions/feats.ark,transcriptions/feats.scp
VLOG[2] (compute-mfcc-feats[5.5.413~1-c4490]:main():compute-mfcc-feats.cc:182) Processed features for key JonnyGutteridge-20081120-hav-b0041
VLOG[2] (compute-mfcc-feats[5.5.413~1-c4490]:main():compute-mfcc-feats.cc:182) Processed features for key JonnyGutteridge-20081120-hav-b0042
LOG (compute-mfcc-feats[5.5.413~1-c4490]:main():compute-mfcc-feats.cc:185)  Done 2 out of 2 utterances.
add-deltas scp:transcriptions/feats.scp ark:transcriptions/delta-feats.ark
gmm-latgen-faster --word-symbol-table=exp/tri3b/graph/words.txt exp/tri3b/final.mdl exp/tri3b/graph/HCLG.fst ark:transcriptions/delta-feats.ark ark,t:transcriptions/lattices.ark
ERROR (gmm-latgen-faster[5.5.413~1-c4490]:LogLikelihoodZeroBased():decodable-am-diag-gmm.cc:50) Dim mismatch: data dim = 39 vs. model dim = 40

[ Stack-Trace: ]
/home/beomgon/kaldi/src/lib/libkaldi-base.so(kaldi::MessageLogger::LogMessage() const+0xb42) [0x7f00383c8692]

Rudolf A. Braun

unread,
Jul 11, 2019, 10:42:53 AM7/11/19
to kaldi-help
There is a mismatch between the features you extracted and the features your model was trained with. Check that you used the same mfcc.conf for both.

Daniel Povey

unread,
Jul 11, 2019, 11:43:37 AM7/11/19
to kaldi-help
Looks to me like it might be about delta+delta-delta features vs. LDA+MLLT. 
What I would do is start from the decoding script that the voxforge example is using, going to decode.1.log or whatever and starting with that command, and then going back to the step that created the CMVN stats and then back to the one that created the MFCC features.
Dan


--
Go to http://kaldi-asr.org/forums.html find out how to join
---
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 post to this group, send email to kaldi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/9e048fb6-ba5e-471c-9594-4acf2e2d8caa%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
Reply all
Reply to author
Forward
0 new messages