kaid mfc特征_few shot mfcc特征是40维-CSDN博客

本文链接：https://blog.csdn.net/libeicuo8108/article/details/90444294

计算13维度特征

if [ $stage -le 6 ]; then
  for part in call_center_26s; do
    steps/make_mfcc.sh --cmd "$train_cmd" --nj 1 data/$part exp/make_mfcc/$part $mfccdir
    steps/compute_cmvn_stats.sh data/$part exp/make_mfcc/$part $mfccdir
  done
fi

steps/make_mfcc.sh --cmd run.pl --mem 2G --nj 1 data/call_center_26s exp/make_mfcc/call_center_26s mfcc
steps/make_mfcc.sh: moving data/call_center_26s/feats.scp to data/call_center_26s/.backup
utils/validate_data_dir.sh: Successfully validated data-directory data/call_center_26s
steps/make_mfcc.sh: [info]: no segments file exists: assuming wav.scp indexed by utterance.
Succeeded creating MFCC features for call_center_26s
steps/compute_cmvn_stats.sh data/call_center_26s exp/make_mfcc/call_center_26s mfcc
Succeeded creating CMVN stats for call_center_26s

mfcc文件夹下获得文件：cmvn_call_center_26s.ark， cmvn_call_center_26s.scp，

raw_mfcc_call_center_26s.1.ark， raw_mfcc_call_center_26s.1.scp

kaldi-master/src/featbin/copy-feats ark:./mfcc/raw_mfcc_call_center_26s.1.ark ark,t:-|head -n 10
/home/joe/MyProjects/kaldi-master/src/featbin/copy-feats ark:./mfcc/raw_mfcc_call_center_26s.1.ark ark,t:- 
4215481_0000_0000  [
  24.15484 -30.40422 -7.251451 -5.942086 -0.3090379 -8.43939 -12.03806 -11.52326 -1.869475 -9.233513 1.437057 -8.801222 2.097739 
  24.15484 -27.9842 -5.511608 -10.46836 -5.926329 -9.012897 -10.03572 -11.83456 8.611436 2.201151 -8.234953 -3.332107 6.853405 
  24.15484 -27.9842 -10.57297 -8.716251 -4.990114 -4.424834 -8.433843 -6.54252 14.10144 10.7261 -1.506598 6.713466 1.636926 
  23.23705 -27.37919 -12.62915 -16.47397 6.244469 3.604277 -11.43735 1.862477 -1.869475 -10.13625 10.4042 3.038257 -9.192201 
  23.23705 -27.07669 -7.725954 -1.999851 -4.053898 -13.60096 -0.4244785 -12.14585 -3.865839 8.520308 -5.431472 -8.340326 -0.2063303 
  22.93112 -30.70673 -6.144278 1.066332 -5.302185 -3.851326 2.779266 -4.052151 -3.366748 -11.88321 -13.56157 -10.87525 8.887736 
  22.62519 -30.40422 -11.04747 -17.68056 -17.06213 -28.04034 -13.63993 -2.495669 -5.036245 -5.47211 -13.98209 -12.94928 -6.427318 
  23.54298 -31.31174 -12.62915 -9.884321 -1.557324 1.597 -9.234778 -13.83973 -13.47521 -4.118006 -5.291298 -11.1057 -8.15537 
  24.15484 -27.68169 -6.776949 -5.50406 -4.36597 2.744015 -8.033374 -6.231224 -10.37518 -18.28875 -3.469034 -0.146925 -6.657724

可以看到mfcc特征是13维度的。下面我们使用add-deltas将维度提升到39维。

查看gmm-latfen-faster解码过程：

gmm-latgen-faster --max-active=7000 --beam=13.0 --lattice-beam=6.0 
--acoustic-scale=0.083333 --allow-partial=true 
--word-symbol-table=exp/tri1/graph_nosp_tgsmall/words.txt 
exp/tri1/final.mdl 
exp/tri1/graph_nosp_tgsmall/HCLG.fst 
"ark,s,cs:apply-cmvn  --utt2spk=ark:data/call_center_26s/split1/1/utt2spk scp:data/call_center_26s/split1/1/cmvn.scp scp:data/call_center_26s/split1/1/feats.scp ark:- | add-deltas  ark:- ark:- |" 
"ark:|gzip -c > exp/tri1/decode_nosp_tgsmall_call_center_26s/lat.1.gz"

./apply-cmvn --utt2spk=ark:data/call_center_26s/split1/1/utt2spk scp:data/call_center_26s/split1/1/cmvn.scp scp:data/call_center_26s/split1/1/feats.scp ark:- | ./add-deltas ark:- ark:cmvn.ark

获得特征 cmvn.ark，可以看出特征是39维，可以用于之后的gmm-latgen-faster 解码

joe@wafer:~/MyProjects/kaldi-master/egs/librispeech/s5$ ~/MyProjects/kaldi-master/src/featbin/copy-feats ark:cmvn.ark ark,t:-|head -n 4
/home/joe/MyProjects/kaldi-master/src/featbin/copy-feats ark:cmvn.ark ark,t:- 
4215481_0000_0000  [
  -41.57018 -30.18891 7.928351 2.303769 26.27834 -8.04239 -2.28777 -18.10191 15.35848 -8.468905 8.756104 -9.345196 5.507133 0 0.7260089 -0.4903194 -1.00746 -1.497945 0.7455605 0.9210769 0.9650178 4.242274 5.135388 -1.555932 3.649849 0.3834037 -0.07342362 0.181502 -0.3368969 -0.1102769 0.2902271 0.3383697 0.4445197 0.572785 -0.3393819 0.4162529 0.4413885 0.4283972 -0.7385951 
  -41.57018 -27.76888 9.668195 -2.222501 20.66105 -8.615897 -0.285429 -18.4132 25.83939 2.965759 -0.915906 -3.876081 10.2628 -0.1835585 0.8470101 -1.407691 -2.383794 0.8425937 2.810189 0.4805619 3.175221 1.597091 1.815414 1.499063 3.919364 -2.30407 -0.09483886 -0.1875522 -0.06959358 0.8963009 0.4649873 -0.005735204 0.7188404 0.2396981 -1.826673 -1.346713 0.2999325 -1.11363 -0.3905961 
  -41.57018 -27.76888 4.606833 -0.470396 21.59726 -4.027833 1.316443 -13.12117 31.32939 11.4907 5.812449 6.169491 5.04632 -0.2753382 0.7260079 -0.8066546 0.1878854 0.4681077 0.229403 2.182552 1.245185 -1.447364 2.317024 0.4902095 0.7292155 -2.065374 -0.08260167 -0.4386302 0.3653672 0.7299631 -0.4766388 -1.512393 0.1802107 -0.3673296 -2.222853 -2.195232 -0.962836 -2.488087 0.2150093

hire特征是40维

pitch_hire特征是44维