Closed ben-8878 closed 4 years ago
Recently, I do a large number of experiments and have found that it may miss utt when using "steps/make_mfcc_pitch". That drives me to reopen the problem,maybe its a bug.
when i used "steps/make_mfcc.sh", that nerver happend.
anyone can answer me?
You'll have to fid more details about where exactly those utterances go missing, e.g. are they in those scp files like data/make_mfcc/train/wav_train.1.scp ? Are they are in (the data-dir)/splitN/1/utt2spk?
i use "fix_data_dir.sh" so it not in "data/make_mfcc/train/wav_train.1.scp" and "splitN/1/utt2spk" ; it just exists in ".backup/wav.scp" ".backup/utt2spk"
i have checked and find the main reason that making mfcc and pitch failtrue is that pitch feature have been extracted failed. first time making mfcc and pitch failtrue, but maybe second time it success the situation of miss utt maybe happen when wav numbers are large
You'll need to do some more debugging yourself to figure out what went wrong.
part log is as follows:
VLOG[2] (compute-mfcc-feats[5.5.546~1-bf0ee]:main():compute-mfcc-feats.cc:182) Processed features for key T36424G00021S1006
WARNING (paste-feats[5.5.546~1-bf0ee]:AppendFeats():paste-feats.cc:45) Length mismatch 715 vs. 712 for utt T36424G00021S1006 exceeds tolerance 2
Possibly you changed the MFCC options or the pitch options and the difference in frame width was enough to cause a mismatch. You might have to either change the tolerance or reduce the difference in frame width.
On Fri, Apr 3, 2020 at 10:59 AM v-yunbin notifications@github.com wrote:
part log is as follows: VLOG[2] (compute-mfcc-feats[5.5.5461-bf0ee]:main():compute-mfcc-feats.cc:182) Processed features for key T36424G00021S1006 WARNING (paste-feats[5.5.5461-bf0ee]:AppendFeats():paste-feats.cc:45) Length mismatch 715 vs. 712 for utt T36424G00021S1006 exceeds tolerance 2
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/kaldi-asr/kaldi/issues/3748#issuecomment-608202637, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAZFLOY55ZDALXCHDQWD4FLRKVGJTANCNFSM4JTTIKMQ .
I get it, thank danpovey‘s reply
@v-yunbin have you figured out the problem?
i have checked and find the main reason that making mfcc and pitch failtrue is that pitch feature have been extracted failed. first time making mfcc and pitch failtrue, but maybe second time it success the situation of miss utt maybe happen when wav numbers are large
Hello, @v-yunbin are you sure that the reason for the extraction failure is that wav numbers are large? I encounter this problemthis problem recently, so I want to ask how you solved this problem。Thank you
If you change the default configure of MFCC_PITCH feature extraction, remember write down your new feature configure in conf/pitch.conf at the sametime. DO NOT only write them in mfcc.conf.
Hello, can you tell me how you solved this problem? I also encountered a similar problem, exactly the same as yours, and it bothered me very much.
Dear maintenance staff:
When i first use "steps/make_mfcc_pitch" to extract mfcc and pitch feature, it works well. But I second use "steps/make_mfcc_pitch" to extract mfcc and pitch feature on same data, 20 percent of data's feature were extracted failed. I think it maybe a bug? but don't know the reason, detail info is as follows: