Extending Gentle Aligner

Week 4

Automate Decoding Process

• Data preparation:

  utt2spk, spk2utt and wav.scp are generated using the script ‘decoding.py’ Find the code here

• Decoding steps:

    compute-mfcc
    ./src/featbin/compute-mfcc-feats --config=voxforge_ru/conf/mfcc.conf scp,p: data/wav.scp ark,scp: data/feats.ark,data/feats.scp
      
    compute-cmvn
    ./src/featbin/compute-cmvn-stats --spk2utt=ark:data/spk2utt scp:data/feats.scp ark,scp:data/cmvn.ark,data/cmvn.scp
  
    gmm-latgen-faster
    gmm-latgen-faster$thread_string --max-active=$max_active --beam=$beam --lattice-beam=$lattice_beam --acoustic-scale=$acwt 
    --allow-partial=true --word-symbol-table=$graphdir/words.txt $decode_extra_opts $model $graphdir/HCLG.fst "$feats" "ark:|gzip -c > 
    $dir/lat.JOB.gz" "ark,t:$dir/words.JOB" "ark,t:$dir/alignments.JOB" || exit 1;
  

• Alignment generation:

    Phoneme Alignment:
    ./src/bin/ali-to-phones --ctm-output exp/tri2a/final.mdl ark: exp/tri2a/decode/alignments.1 1.ctm
  
    Word-level Alignment:
    ./steps/get_ctm.sh ./transcript ./data/lang ./exp/tri2a/decode/
  

• Preparing data for visualization:

    [link to code for converting to json format](https://github.com/shreya2111/gentle-labs/tree/master/v1/scripts)
  

• Next: Building a Russian language model - Reverse engineering HCLG.fst Week 5
• Prev: Visualization using Javascript Week 3
• main page

Tools: Kaldi, Python, C, Bash Scripting

Link to GSoC Project Repository