Extending Gentle Aligner

Week 6

Generating a Language Model

• Preparing for integration with Gentle Aligner
  Acoustic model:

   * Pick final.mdl and tree for the langauge that you want to build langauge model for.
   * Place final.mdl inside data/tdnn_7b_chain_online/
   * Place tree inside data/tdnn_7b_chain_online/
  

  Lexicon: * Place words.txt inside data/tdnn_7b_chain_online/graph_pp/words.txt

• Did we get HCLG.fst?

  run script generateLM.py with three arguments, 1. path to the utterance.txt 2. proto_dir: path to the data directory containing langdir and tdnn_7b_chain_online 3. path toa pristine copy of Kaldi

  Ex: python3 generateLM.py path-to/utterance.txt path-to/data This script calls langauge_model functions of Gentle which generate a bigram grammar G.fst, and compiles L.fst, C.fst and G.fst into HCLG.fst.

    Inputs taken by Gentle are:
       langdir/L.fst
       langdir/L_disambig.fst
       langdir/phones/disambig.int
       tdnn_7b_chain_online/final.mdl
       tdnn_7b_chain_online/tree
       tdnn_7b_chain_online/graph_pp/words.txt
  

• After this step you would have generated temp_HCLG.fst which gets saved as tdnn_7b_chain_online/graph_pp/HCLG.fst. This HCLG.fst will be used in the next step of decoding utterances and producing their word & phoneme level alignments.

• Next: Decoding using an utterance-level-customized langauge model Week 7
• Prev: Groundwork for Generating a Language Model Week 5
• main page

Tools: Kaldi, Python, C, Bash Scripting

Link to GSoC Project Repository