search

alumae / kaldi-offline-transcriber

Offline transcription system for Estonian using Kaldi
Other
226 stars 55 forks source link

transcription of sample intervjuu201306211256.mp3 fails #2

Closed archibaldhaddock closed 9 years ago

archibaldhaddock commented 10 years ago

Hello,

I've just tried to decode intervjuu201306211256.mp3.

Everything seems to work fine until the "final pass of acoustic scoring". I've got this message :

steps/decode_nnet.sh: missing file build/trans/intervjuu201306211256/nnet5c1_pruned/final.nnet make: *\ [build/trans/intervjuu201306211256/nnet5c1_pruned/decode/log] Erreur 1

The complete output of the command "make build/output/intervjuu201306211256.txt" can be found here : http://wikisend.com/download/446498/make.interview.out

And the make .init output is there : http://wikisend.com/download/376740/make.init.out

Thanks a lot !

alumae commented 10 years ago

Have you modified the Makefile? In Makefile, we execute 'steps/decode_nnet_cpu.sh', but you seem to execute 'steps/decode_nnet_cpu.sh'

archibaldhaddock commented 10 years ago

Hello,

Thank you for your quick answer ! Yes indeed, I have changed steps/decode_nnet_cpu.sh to steps/decode_nnet.sh in Makefile because I first tried to run the test with kaldi-stable that has no decode_nnet.sh script... Well, by now, I 've changed it back and relaunch.

Now there is a message like "cannot find nnet-latgen-faster"

steps/decode_nnet_cpu.sh: feature type is lda steps/decode_nnet_cpu.sh: using transforms from build/trans/intervjuu201306211256/tri3b_mmi_pruned/decode run.pl: job failed, log is in build/trans/intervjuu201306211256/nnet5c1_pruned/decode/log/decode.1.log make: *\ [build/trans/intervjuu201306211256/nnet5c1_pruned/decode/log] Erreur 1

cat build/trans/intervjuu201306211256/nnet5c1_pruned/decode/log/decode.1.log

nnet-latgen-faster --max-active=7000 --beam=15.0 --lattice-beam=8.0 --acoustic-scale=0.1 --allow-partial=true --word-symbol-table=build/fst/tri3b/graph_prunedlm/words.txt build/trans/intervjuu201306211256/nnet5c1_pruned/final.mdl build/fst/tri3b/graph_prunedlm/HCLG.fst "ark,s,cs:apply-cmvn --norm-vars=false --utt2spk=ark:build/trans/intervjuu201306211256/split1/1/utt2spk scp:build/trans/intervjuu201306211256/split1/1/cmvn.scp scp:build/trans/intervjuu201306211256/split1/1/feats.scp ark:- | splice-feats ark:- ark:- | transform-feats build/trans/intervjuu201306211256/nnet5c1_pruned/final.mat ark:- ark:- | transform-feats --utt2spk=ark:build/trans/intervjuu201306211256/split1/1/utt2spk ark,s,cs:build/trans/intervjuu201306211256/tri3b_mmi_pruned/decode/trans.1 ark:- ark:- |" "ark:|gzip -c > build/trans/intervjuu201306211256/nnet5c1_pruned/decode/lat.1.gz" Started at mercredi 5 février 2014, 18:10:37 (UTC+0100)

bash: line 1: nnet-latgen-faster : commande introuvable Accounting: time=0 threads=1

I've just checked that $KALDI_ROOT is OK

archibaldhaddock commented 10 years ago

And if I put the complete path of nnet-letgen-faster in decode_nnet_cpu.sh, the processing ends up with : steps/decode_nnet_cpu.sh: feature type is lda steps/decode_nnet_cpu.sh: using transforms from build/trans/intervjuu201306211256/tri3b_mmi_pruned/decode run.pl: job failed, log is in build/trans/intervjuu201306211256/nnet5c1_pruned/decode/log/decode.1.log

decode1.log is

/home/gwen/dev/projet_voxilab/kaldi-trunk/src/nnet2bin/nnet-latgen-faster --max-active=7000 --beam=15.0 --lattice-beam=8.0 --acoustic-scale=0.1 --allow-partial=true --word-symbol-table=build/fst/tri3b/graph_prunedlm/words.txt build/trans/intervjuu201306211256/nnet5c1_pruned/final.mdl build/fst/tri3b/graph_prunedlm/HCLG.fst "ark,s,cs:apply-cmvn --norm-vars=false --utt2spk=ark:build/trans/intervjuu201306211256/split1/1/utt2spk scp:build/trans/intervjuu201306211256/split1/1/cmvn.scp scp:build/trans/intervjuu201306211256/split1/1/feats.scp ark:- | splice-feats ark:- ark:- | transform-feats build/trans/intervjuu201306211256/nnet5c1_pruned/final.mat ark:- ark:- | transform-feats --utt2spk=ark:build/trans/intervjuu201306211256/split1/1/utt2spk ark,s,cs:build/trans/intervjuu201306211256/tri3b_mmi_pruned/decode/trans.1 ark:- ark:- |" "ark:|gzip -c > build/trans/intervjuu201306211256/nnet5c1_pruned/decode/lat.1.gz" Started at mercredi 5 février 2014, 20:45:35 (UTC+0100)

/home/gwen/dev/projet_voxilab/kaldi-trunk/src/nnet2bin/nnet-latgen-faster --max-active=7000 --beam=15.0 --lattice-beam=8.0 --acoustic-scale=0.1 --allow-partial=true --word-symbol-table=build/fst/tri3b/graph_prunedlm/words.txt build/trans/intervjuu201306211256/nnet5c1_pruned/final.mdl build/fst/tri3b/graph_prunedlm/HCLG.fst 'ark,s,cs:apply-cmvn --norm-vars=false --utt2spk=ark:build/trans/intervjuu201306211256/split1/1/utt2spk scp:build/trans/intervjuu201306211256/split1/1/cmvn.scp scp:build/trans/intervjuu201306211256/split1/1/feats.scp ark:- | splice-feats ark:- ark:- | transform-feats build/trans/intervjuu201306211256/nnet5c1_pruned/final.mat ark:- ark:- | transform-feats --utt2spk=ark:build/trans/intervjuu201306211256/split1/1/utt2spk ark,s,cs:build/trans/intervjuu201306211256/tri3b_mmi_pruned/decode/trans.1 ark:- ark:- |' 'ark:|gzip -c > build/trans/intervjuu201306211256/nnet5c1_pruned/decode/lat.1.gz' ERROR (nnet-latgen-faster:ExpectOneOrTwoTokens():nnet-component.cc:143) Expecting token FixedLinearComponent or CuMatrix but got Matrix ERROR (nnet-latgen-faster:ExpectOneOrTwoTokens():nnet-component.cc:143) Expecting token FixedLinearComponent or CuMatrix but got Matrix

[stack trace: ] kaldi::KaldiGetStackTrace() kaldi::KaldiErrorMessage::~KaldiErrorMessage() /home/gwen/dev/projet_voxilab/kaldi-trunk/src/nnet2bin/nnet-latgen-faster() [0x6401d9] kaldi::nnet2::FixedLinearComponent::Read(std::istream&, bool) kaldi::nnet2::Component::ReadNew(std::istream&, bool) kaldi::nnet2::Nnet::Read(std::istream&, bool) kaldi::nnet2::AmNnet::Read(std::istream&, bool) /home/gwen/dev/projet_voxilab/kaldi-trunk/src/nnet2bin/nnet-latgen-faster(main+0x611) [0x5f0b0e] /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf5) [0x2b04e4379995] /home/gwen/dev/projet_voxilab/kaldi-trunk/src/nnet2bin/nnet-latgen-faster() [0x5f0434]

Accounting: time=0 threads=1

alumae commented 10 years ago

So which SVN branch are you now using? In the latest trunk, nnet-latgen-faster should exist under $KALDI_ROOT/src/nnet-cpubin which is added to the PATH in the Makefile.

archibaldhaddock commented 10 years ago

You are right again, I don't know how it comes but my svn revision was 3460 (I downloaded it on monday)...I have updated it and I have now the same error than when I had manually changed nnet-latgen-faster 's path. It seems to have some relations with cuda matrices (see previous post). kaldi was generated with ./configure --use-cuda=yes...I'll try to fix it myself... Thanks again for your help !

alumae commented 10 years ago

I'm actually using Kaldi trunk revision 2720 with this system. I'll check if it still works for me with the latest Kaldi revision.

alumae commented 10 years ago

I reproduced your error with the latest Kaldi. I hope to fix it soon.

archibaldhaddock commented 10 years ago

Ok. Thank you for your work!

alumae commented 10 years ago

It seems that the current Kaldi trunk is incompatible with the neural network acoustic models that I have (they were created last summer). I plan to retrain the models soon using the new Kaldi code, which should solve the problem, but it may take some time (weeks).

However, you should be able to run the system with Kaldi's revision 2720 if you really need it. I'm not sure about the actual revision that broke it but this is the revision that I'm using in production with this system.

archibaldhaddock commented 10 years ago

Hello,

I am sorry to use this thread to contact you but I could not find another way to do it.

I had a look at kaldi's code and did a little profiling around, testing the first steps of the processing of your sample file. I am now planning to optimize it.

The company (http://www.melvenn.com/) I am creating with a friend is doing this kind of work (amongst other things) : improving the performance of open sources projects.

Do you know people using kaldi that could be interested in a little (big?) speedup of its computing performance and would be ready to pay for that ?

Thanks a lot if you ever answer this somewhat peculiar question ;-) !