Online Decoding for chain-tdnn

[sameer khurana]

Hi,

I have been trying to use online decoding using the chain-tdnn model that I trained. The decode.* log is empty for almost all the utterances. See below:

decode.1.log:

The command:

online2-wav-nnet3-latgen-faster --online=true --do-endpointing=false --frame-subsampling-factor=3 --config=exp/mer80/chain/tdnn_7b_online_1/conf/online.conf --max-active=7000 --beam=15.0 --lattice-beam=6.0 --acoustic-scale=0.1 --word-symbol-table=exp/mer80/chain/tdnn_7b/graph/words.txt exp/mer80/chain/tdnn_7b_online_1/final.mdl exp/mer80/chain/tdnn_7b/graph/HCLG.fst ark:data/dev_non_overlap_hires/split100/1/spk2utt "ark,s,cs:extract-segments scp,p:data/dev_non_overlap_hires/split100/1/wav.scp data/dev_non_overlap_hires/split100/1/segments ark:- |" "ark:|gzip -c >exp/mer80/chain/tdnn_7b_online_1/decode_non_overlap/lat.1.gz"

Continuation of log:

LOG (online2-wav-nnet3-latgen-faster:ComputeDerivedVars():ivector-extractor.cc:183) Computing derived variables for iVector extractor

LOG (online2-wav-nnet3-latgen-faster:ComputeDerivedVars():ivector-extractor.cc:204) Done.

extract-segments scp,p:data/dev_non_overlap_hires/split100/1/wav.scp data/dev_non_overlap_hires/split100/1/segments ark:-

01BE8E7B-C179-42E3-8521-109C2C732334_spk-0001_seg-0001345:0001695

LOG (online2-wav-nnet3-latgen-faster:main():online2-wav-nnet3-latgen-faster.cc:272) Decoded utterance 01BE8E7B-C179-42E3-8521-109C2C732334_spk-00\

01_seg-0001345:0001695

01BE8E7B-C179-42E3-8521-109C2C732334_spk-0001_seg-0001695:0002182 nslT xlAlhA AlDw'

LOG (online2-wav-nnet3-latgen-faster:main():online2-wav-nnet3-latgen-faster.cc:272) Decoded utterance 01BE8E7B-C179-42E3-8521-109C2C732334_spk-00\

01_seg-0001695:0002182

01BE8E7B-C179-42E3-8521-109C2C732334_spk-0001_seg-0002182:0002565

LOG (online2-wav-nnet3-latgen-faster:main():online2-wav-nnet3-latgen-faster.cc:272) Decoded utterance 01BE8E7B-C179-42E3-8521-109C2C732334_spk-00\

01_seg-0002182:0002565

01BE8E7B-C179-42E3-8521-109C2C732334_spk-0001_seg-0003154:0003720

LOG (online2-wav-nnet3-latgen-faster:main():online2-wav-nnet3-latgen-faster.cc:272) Decoded utterance 01BE8E7B-C179-42E3-8521-109C2C732334_spk-00\

01_seg-0003154:0003720

As you can see, empty decoded output.

I performed the below two steps in an attempt at online decoding:

•  steps/online/nnet3/prepare_online_decoding.sh --mfcc-config conf/mfcc_hires.conf data/lang exp/mer80/nnet3/extractor exp/mer80/chain/tdnn_7b exp/mer80/chain/tdnn_7b_online

•  steps/online/nnet3/decode.sh --cmd "hostd.pl" --nj 100 exp/mer80/chain/tdnn_7b/graph data/dev_non_overlap_hires exp/mer80/chain/tdnn_7b_online/decode_non_overlap
  

Offline decoding works fine.

I also attempted online decoding using NNET3 TDNN model, which worked fine. The decode.1.log is given below:

# online2-wav-nnet3-latgen-faster --online=true --do-endpointin\

g=false --config=exp/mer80/nnet3/nnet_tdnn_a_online/conf/online.conf --max-active=7000 --beam=15.0 --lattice-beam=6.0 --acoustic-scale=0.1 --word\

-symbol-table=exp/mer80/tri4/graph//words.txt exp/mer80/nnet3/nnet_tdnn_a_online/final.mdl exp/mer80/tri4/graph//HCLG.fst ark:exp/mer80/nnet3/nne\

t_tdnn_a_online/decode_non_overlap_utt/per_utt/utt2spk.1 "ark,s,cs:extract-segments scp,p:data/dev_non_overlap_hires/split100/1/wav.scp data/dev_\

non_overlap_hires/split100/1/segments ark:- |" "ark:|gzip -c >exp/mer80/nnet3/nnet_tdnn_a_online/decode_non_overlap_utt/lat.1.gz"

/data/sls/qcri-scratch/sameer/language_modelling/kaldi_noGPU/kaldi/src/online2bin/online2-wav-nnet3-latgen-faster --online=true --do-endpointing=\

false --config=exp/mer80/nnet3/nnet_tdnn_a_online/conf/online.conf --max-active=7000 --beam=15.0 --lattice-beam=6.0 --acoustic-scale=0.1 --word-s\

ymbol-table=exp/mer80/tri4/graph//words.txt exp/mer80/nnet3/nnet_tdnn_a_online/final.mdl exp/mer80/tri4/graph//HCLG.fst ark:exp/mer80/nnet3/nnet_\

tdnn_a_online/decode_non_overlap_utt/per_utt/utt2spk.1 'ark,s,cs:extract-segments scp,p:data/dev_non_overlap_hires/split100/1/wav.scp data/dev_no\

n_overlap_hires/split100/1/segments ark:- |' 'ark:|gzip -c >exp/mer80/nnet3/nnet_tdnn_a_online/decode_non_overlap_utt/lat.1.gz'

LOG (online2-wav-nnet3-latgen-faster:ComputeDerivedVars():ivector-extractor.cc:183) Computing derived variables for iVector extractor

LOG (online2-wav-nnet3-latgen-faster:ComputeDerivedVars():ivector-extractor.cc:204) Done.

extract-segments scp,p:data/dev_non_overlap_hires/split100/1/wav.scp data/dev_non_overlap_hires/split100/1/segments ark:-

01BE8E7B-C179-42E3-8521-109C2C732334_spk-0001_seg-0001345:0001695 >hlA bkm fy h*h AlHlqp mn brnAmj AlwAqE AlErby

LOG (online2-wav-nnet3-latgen-faster:main():online2-wav-nnet3-latgen-faster.cc:272) Decoded utterance 01BE8E7B-C179-42E3-8521-109C2C732334_spk-00\

01_seg-0001345:0001695

01BE8E7B-C179-42E3-8521-109C2C732334_spk-0001_seg-0001695:0002182 wAlty nslT xlAlhA AlDw' ElY dwr AlAntxAbAt AlmHlyp Almgrbyp

LOG (online2-wav-nnet3-latgen-faster:main():online2-wav-nnet3-latgen-faster.cc:272) Decoded utterance 01BE8E7B-C179-42E3-8521-109C2C732334_spk-00\

01_seg-0001695:0002182

01BE8E7B-C179-42E3-8521-109C2C732334_spk-0001_seg-0002182:0002565 fy tEzyz msAr Al<SlAH AlsyAsy fy Almmlkp

LOG (online2-wav-nnet3-latgen-faster:main():online2-wav-nnet3-latgen-faster.cc:272) Decoded utterance 01BE8E7B-C179-42E3-8521-109C2C732334_spk-00\

01_seg-0002182:0002565

01BE8E7B-C179-42E3-8521-109C2C732334_spk-0001_seg-0003154:0003720 msAr Al<SlAH AlsyAsy fy Almgrb ElY mHk AlAntxAbAt AlmHlyp

LOG (online2-wav-nnet3-latgen-faster:main():online2-wav-nnet3-latgen-faster.cc:272) Decoded utterance 01BE8E7B-C179-42E3-8521-109C2C732334_spk-00\

01_seg-0003154:0003720

01BE8E7B-C179-42E3-8521-109C2C732334_spk-0001_seg-0003720:0004181 mnETfA yjmE EAmp AlmgArbp ElY >hmyth

LOG (online2-wav-nnet3-latgen-faster:main():online2-wav-nnet3-latgen-faster.cc:272) Decoded utterance 01BE8E7B-C179-42E3-8521-109C2C732334_spk-00\

01_seg-0003720:0004181

01BE8E7B-C179-42E3-8521-109C2C732334_spk-0001_seg-0004181:0004667 wqd >qblwA ElY rEt AqtrAE stnEks ntA}jh

LOG (online2-wav-nnet3-latgen-faster:main():online2-wav-nnet3-latgen-faster.cc:272) Decoded utterance 01BE8E7B-C179-42E3-8521-109C2C732334_spk-00\

01_seg-0004181:0004667

Any suggestions?

[Daniel Povey]

You need to use --acoustic-scale=1.0 for chain models.
Dan

> --
> 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.
> For more options, visit https://groups.google.com/d/optout.