Missing utt xx from input scp:data/dev/mcep_feats.scp, Letter not in LTS tree

[ikhfa]

I'm running the tts_dnn_arctic sample with 48K version. Is it normal getting this kind of warning?

WARNING (paste-feats:main():paste-feats.cc:136) Missing utt slt_arctic_b0339 from input scp:data/train/mcep_feats.scp ........... WARNING (paste-feats:main():paste-feats.cc:136) Missing utt slt_arctic_b0539 from input scp:data/train/mcep_feats.scp LOG (paste-feats:main():paste-feats.cc:157) Done 0 utts, errors on 1132 steps/compute_cmvn_stats.sh data/train exp/compute_cmvn/train data/train Error computing CMVN stats

steps/make_pitch.sh --pitch-config conf/pitch-48k.conf data/dev exp/make_pitch/dev /tmp/dnn_feats/arctic2 utils/validate_data_dir.sh: WARNING: you have only one speaker. This probably a bad idea. Search for the word 'bold' in http://kaldi.sourceforge.net/data_prep.html for more information. utils/validate_data_dir.sh: Successfully validated data-directory data/dev steps/make_pitch.sh: [info]: no segments file exists: assuming wav.scp indexed by utterance. Succeeded creating Pitch features for dev steps/compute_cmvn_stats.sh data/dev exp/compute_cmvn_pitch/dev /tmp/dnn_feats/arctic2 Succeeded creating CMVN stats for dev copy-feats 'scp:awk -v spk=slt '\''($1 == spk){print}'\'' data/dev/cmvn.scp |' ark,t:- LOG (copy-feats:main():copy-feats.cc:100) Copied 1 feature matrices. 129.726 using wsizes: 35 18 Retained 100000 utterances per speaker from data-dir data/dev and put it in data/dev_slt, reducing #utt from 60 to 60 steps/make_pitch.sh --pitch-config conf/pitch-48k.conf --frame_length 18 data/dev_slt exp/make_pitch/dev_slt /tmp/dnn_feats/arctic2 utils/validate_data_dir.sh: WARNING: you have only one speaker. This probably a bad idea. Search for the word 'bold' in http://kaldi.sourceforge.net/data_prep.html for more information. utils/validate_data_dir.sh: Successfully validated data-directory data/dev_slt steps/make_pitch.sh: [info]: no segments file exists: assuming wav.scp indexed by utterance. Succeeded creating Pitch features for dev_slt steps/make_bndap.sh --bndap-config conf/bndap-48k.conf --frame_length 35 data/dev_slt exp/make_bndap/dev_slt /tmp/dnn_feats/arctic2 utils/validate_data_dir.sh: WARNING: you have only one speaker. This probably a bad idea. Search for the word 'bold' in http://kaldi.sourceforge.net/data_prep.html for more information. utils/validate_data_dir.sh: Successfully validated data-directory data/dev_slt steps/make_bndap.sh: [info]: no segments file exists: assuming wav.scp indexed by utterance. Succeeded creating Bndap features for dev_slt steps/make_mcep.sh --sample-frequency 48000 --frame_length 18 data/dev_slt exp/make_mcep/dev_slt /tmp/dnn_feats/arctic2 utils/validate_data_dir.sh: WARNING: you have only one speaker. This probably a bad idea. Search for the word 'bold' in http://kaldi.sourceforge.net/data_prep.html for more information. utils/validate_data_dir.sh: Successfully validated data-directory data/dev_slt steps/make_mcep.sh: [info]: no segments file exists: assuming wav.scp indexed by utterance. run.pl: 4 / 4 failed, log is in exp/make_mcep/dev_slt/makemcep..log cat: data/dev__/mcep_feats.scp: No such file or directory paste-feats --length-tolerance=1 scp:data/dev/pitch_feats.scp scp:data/dev/mcep_feats.scp scp:data/dev/bndap_feats.scp ark,scp:/tmp/dnn_feats/arctic2/dev_cmp_feats.ark,data/dev/feats.scp WARNING (paste-feats:main():paste-feats.cc:136) Missing utt slt_arctic_a0002 from input scp:data/dev/mcep_feats.scp .......... WARNING (paste-feats:main():paste-feats.cc:136) Missing utt slt_arctic_a0592 from input scp:data/dev/mcep_feats.scp LOG (paste-feats:main():paste-feats.cc:157) Done 0 utts, errors on 60 steps/compute_cmvn_stats.sh data/dev exp/compute_cmvn/dev data/dev Error computing CMVN stats

Step 2: label creation

idlaktxp --pretty --tpdb=/home/osboxes/Tesis/idlak/egs/tts_dnn_arctic/s1/../../../idlak-data/en/ga/ data/train/text.xml data/train/text_norm.xml WARNING (idlaktxp:GetPron():txplts.cc:96) Letter not in LTS tree: ' .............. WARNING (idlaktxp:GetPron():txplts.cc:96) Letter not in LTS tree: ' idlaktxp --pretty --tpdb=/home/osboxes/Tesis/idlak/egs/tts_dnn_arctic/s1/../../../idlak-data/en/ga/ data/dev/text.xml data/dev/text_norm.xml WARNING (idlaktxp:GetPron():txplts.cc:96) Letter not in LTS tree: ' idlaktxp --pretty --tpdb=/home/osboxes/Tesis/idlak/egs/tts_dnn_arctic/s1/../../../idlak-data/en/ga/ data/full/text.xml data/full/text_norm.xml WARNING (idlaktxp:GetPron():txplts.cc:96) Letter not in LTS tree: ' WARNING (idlaktxp:GetPron():txplts.cc:96) Letter not in LTS tree: '

[bpotard]

Most of the warnings I see are fine, however:

run.pl: 4 / 4 failed, log is in exp/make_mcep/dev_slt/make_mcep..log

Means that the mcep extraction failed. I suspect it is because sptk was not correctly build when running make in "tools". Have you build all the tools successfully, i.e. did make in the tools/ directory succeed? If it succeeded, you can look at exp/make_mcep/dev_slt/make_mcep.1.log to see what went wrong. In any case, idlak cannot build the voice without mcep features.

[ikhfa]

I think there's nothing wrong with the 'tools' build. But, I'll check it again. This is the content of the log:

copy-feats --compress=false "ark,s,cs:compute-mcep-feats.sh --frame_length 18 --srate 48000 scp,p:exp/make_mcep/dev_slt/wav_dev_slt.1.scp ark:- |" ark,scp:/tmp/dnn_feats/arctic2/raw_mcep_dev_slt.1.ark,/tmp/dnn_feats/arctic2/raw_mcep_dev_slt.1.scp

Started at Wed Aug 10 17:41:51 WIB 2016

copy-feats --compress=false 'ark,s,cs:compute-mcep-feats.sh --frame_length 18 --srate 48000 scp,p:exp/make_mcep/dev_slt/wav_dev_slt.1.scp ark:- |' ark,scp:/tmp/dnn_feats/arctic2/raw_mcep_dev_slt.1.ark,/tmp/dnn_feats/arctic2/raw_mcep_dev_slt.1.scp

/.../idlak/egs/tts_dnn_arctic/s1/../../../src/featbin//compute-mcep-feats.sh: 34: /.../idlak/egs/tts_dnn_arctic/s1/utils//parse_options.sh: Syntax error: Bad for loop variable

WARNING (copy-feats:Close():kaldi-io.cc:446) Pipe compute-mcep-feats.sh --frame_length 18 --srate 48000 scp,p:exp/make_mcep/dev_slt/wav_dev_slt.1.scp ark:- | had nonzero return status 512

LOG (copy-feats:main():copy-feats.cc:100) Copied 0 feature matrices.

Ended (code 256) at Wed Aug 10 17:41:51 WIB 2016, elapsed time 0 seconds

[bpotard]

Right, I see what the problem is.

In src/featbin/compute-mcep-feats.sh, the shell to use should be bash (#!/bin/bash), not sh. It is not a problem on systems where sh is just a symlink to bash, but it is not your case obviously!

I have now fixed it in the repository.

[bpotard]

Is that now working for you? You can either manually modify the first line of src/featbin/compute-mcep-feats.sh or you can perform a git pull.

[ikhfa]

Yes, it is working now (took me almost 7 hours to train the DNN :D). Thank you. No errors reported, except several warnings like this:

KALDICEX[Thu Aug 11 14:16:51 2016] INFO:Started Logging make-fullctx-ali-dnn --max-sil-phone=15 exp-align/quin_ali_full/final.mdl 'ark:gunzip -c exp-align/quin_ali_full/ali.*.gz|' ark,t:data/full/cex.ark ark,t:data/full/ali WARNING (make-fullctx-ali-dnn:main():make-fullctx-ali-dnn.cc:239) Merge of alignment and contexts failed for key slt_arctic_a0086 mismatching number of phones contexts:28 alignment:29`

And tons of this:

copy-feats scp:data/dev/feats.scp ark,t:- LOG (copy-feats:main():copy-feats.cc:100) Copied 60 feature matrices. 1 2 3 61 64 25 x2x : warning: input data is over the range of type 'short'! .......... x2x : warning: input data is over the range of type 'short'! sox WARN dither: dither clipped 3 samples; decrease volume?

And then in 5b. Alice sample synthesis:

Traceback (most recent call last): File "/.../idlak/egs/tts_dnn_arctic/s1/../../../idlak-voice-build/utils/idlak_make_lang.py", line 468, in main() File "/.../idlak/egs/tts_dnn_arctic/s1/../../../idlak-voice-build/utils/idlak_make_lang.py", line 463, in main forward_context(logger, args[0], args[1], args[2], opts.root_name) File "/.../idlak/egs/tts_dnn_arctic/s1/../../../idlak-voice-build/utils/idlak_make_lang.py", line 64, in forward_context dom = parse(input_fname) File "/usr/lib/python2.7/xml/dom/minidom.py", line 1918, in parse return expatbuilder.parse(file) File "/usr/lib/python2.7/xml/dom/expatbuilder.py", line 924, in parse result = builder.parseFile(fp) File "/usr/lib/python2.7/xml/dom/expatbuilder.py", line 207, in parseFile parser.Parse(buffer, 0) xml.parsers.expat.ExpatError: not well-formed (invalid token): line 2561, column 82 cat: data/eval/cex.ark: No such file or directory copy-feats ark:- ark,scp:/tmp/dnn_feats/arctic2/in_durfeats_eval.ark,/tmp/dnn_feats/arctic2/in_durfeats_eval.scp LOG (copy-feats:main():copy-feats.cc:100) Copied 0 feature matrices. steps/compute_cmvn_stats.sh lbldurdata/eval lbldurdata/eval lbldurdata/eval Error computing CMVN stats

nnet-forward exp_dnn/tts_dnn_dur_3_delta_quin5/input_final.feature_transform ark:- ark:- LOG (copy-feats:main():copy-feats.cc:100) Copied 0 feature matrices. LOG (nnet-forward:main():nnet-forward.cc:219) Done 0 files in 1.58946e-08min, (fps 0) LOG (apply-cmvn:main():apply-cmvn.cc:143) Applied cepstral mean and variance normalization to 0 utterances, errors on 0 LOG (nnet-forward:main():nnet-forward.cc:219) Done 0 files in 0min, (fps -nan) WARNING (nnet-forward:Close():kaldi-io.cc:446) Pipe copy-feats scp:lbldurdata/eval/feats.scp ark:- | nnet-forward exp_dnn/tts_dnn_dur_3_delta_quin5/input_final.feature_transform ark:- ark:- | apply-> cmvn --print-args=false --norm-vars=true exp_dnn/tts_dnn_dur_3_delta_quin5/cmvn_glob.ark ark:- ark:- | had nonzero return status 256 cat: 'exp_dnn/tts_dnn_dur_3_delta_quin5/tst_forward/cmp/*.cmp': No such file or directory Traceback (most recent call last): File "utils/make_fullctx_mlf_dnn.py", line 112, in main() File "utils/make_fullctx_mlf_dnn.py", line 107, in main make_fullctx_mlf_dnn(args[0], args[1], args[2]) File "utils/make_fullctx_mlf_dnn.py", line 51, in make_fullctx_mlf_dnn flabs = parse_fullctx(fullctx) File "utils/make_fullctx_mlf_dnn.py", line 28, in parse_fullctx fp = open(fullctxfile, "r") IOError: [Errno 2] No such file or directory: 'data/eval/cex.ark' copy-feats ark:data/eval/feat.ark ark,scp:/tmp/dnn_feats/arctic2/in_feats_eval.ark,/tmp/dnn_feats/arctic2/in_feats_eval.scp LOG (copy-feats:main():copy-feats.cc:100) Copied 0 feature matrices. steps/compute_cmvn_stats.sh lbldata/eval lbldata/eval lbldata/eval Error computing CMVN stats

And, lastly:

nnet-forward exp_dnn/tts_dnn_train_3_deltasc2_quin5/input_final.feature_transform ark:- ark:- LOG (copy-feats:main():copy-feats.cc:100) Copied 0 feature matrices. LOG (nnet-forward:main():nnet-forward.cc:219) Done 0 files in 1.58946e-08min, (fps 0) LOG (apply-cmvn:main():apply-cmvn.cc:143) Applied cepstral mean and variance normalization to 0 utterances, errors on 0 LOG (nnet-forward:main():nnet-forward.cc:219) Done 0 files in 0min, (fps -nan) WARNING (nnet-forward:Close():kaldi-io.cc:446) Pipe copy-feats scp:lbldata/eval/feats.scp ark:- | nnet-forward exp_dnn/tts_dnn_train_3_deltasc2_quin5/input_final.feature_transform ark:- ark:- | apply-cmvn --print-args=false --norm-vars=true exp_dnn/tts_dnn_train_3_deltasc2_quin5/cmvn_glob.ark ark:- ark:- | had nonzero return status 256 1 6 7 183 190 75 cat: 'exp_dnn/tts_dnn_train_3_deltasc2_quin5/tst_forward/cmp/*.cmp': No such file or directory utils/mlsa_synthesis_63_mlpg.sh: line 79: $cmp: ambiguous redirect Extracting variances... Creating pdfs... Running mlpg... mcep smoothing utils/mlsa_synthesis_63_mlpg.sh: line 123: $mcep: ambiguous redirect f0 smoothing bndap smoothing utils/mlsa_synthesis_63_mlpg.sh: line 127: $bap: ambiguous redirect utils/mlsa_synthesis_63_mlpg.sh: line 130: $f0: ambiguous redirect utils/mlsa_synthesis_63_mlpg.sh: line 159: $mcep.float: ambiguous redirect Usage: excitation.py [options] [ [

]] e.g. python to_wav test.raw

excitation.py: error: Too many arguments

[bpotard]

Ok, there are several issues there:

1. you are using a recent version of g++ that triggered a bug (that we had not noticed) in idlak txplts; this ended up generating some random garbage in the normalised files in rare occasion, which sometimes caused crashes. This has now been fixed in the repository.
2. I think you may have to remove the dnn features folder entirely (rm -rf /tmp/dnn_feats/) then rerun the training
3. The synthesis does not run at all as the label file has not been generated correctly (because of the bug in 1.); I think I saw in some other logs that you may need to install numpy anyway for the synthesis to work.

I have now added a more thorough check for all required dependencies in tools/; it now checks for sox, python-numpy, csh, and some other things. You should update your git repository, rebuild things in the tools and src folders, probably regenerate the features, train the DNN, and normally everything should now be working fine.

Sorry about these issues!