All the acoustic models are build from the directory "am", the language models are build from the directory "lm".

The steps I took to build the models are:

• Run kaldi_prep.py ar-mgb2016 train am/train and kaldi_prep.py ar-mgb2016 train_wmer0 am/train_wmer0. (kaldi_prep.py comes from https://github.com/psmit/kaldi-prep )
• Train GMM acoustic models with common/train_am.sh
• Train BLSTM DNN with:
  • common/train_chain_dataprep.sh gr
  • common/train_chain_prep.sh gr_tdnn_blstm_9
  • common/train_chain_model.sh gr_tdnn_blstm_9_a
  • run common/train_chain_extra_dataset.sh for dev/test set
• Train vari-gram LM in the lm directory
  • create files mgb2/mgb2.xz, mgb2/devel-mgb2.xz and devel-egy.xz with in the first the lm training data (buckwalter) and in the latter 2 the dev transcriptions for mgb-2 and egy (I can't find anymore the scripts that did this)
  • create some directories, e.g. mgb2/chars, mgb2/morfessor_f2_a0.1_tokens, mgb2/word_f1. These directories contain the parameters in their filename. f2 is a frequency threshold of 2, a0.1 is a morfessor alpha of 0.1.
  • Run common/run_morph_job.sh mgb2/morfessor_f2_a0.1_tokens or common/run_char_job.sh mgb2/chars or common/run_word_job.sh mgb2/word_f1. For Morfessor, I think you also had to run common/run_morphseg_job_dev.sh and common/run_morphseg_job.sh
  • Make directories like varik_aff / varik_pre / varik_suf / varik_wma in the char of morfessor directories and varik_word in the word directory. "aff" is left-right marked, "pre" is left-marked, "suf" is right-marked, "wma" is word boundary marker
  • On these varik* directories, run common/run_prep_dev.sh and common/run_prep_varikn.sh before running common/train_varikn_background.sh. varigram_kn must be on the path (https://github.com/vsiivola/variKN)
• Once a small arpa is created, put it in am/data/lm/MODEL_NAME/small/arpa and put "aff" "word" "suf" "pre" "wma" in am/data/lm/MODEL_NAME/type
• Run from am the command common/process_lm.sh MODEL_NAME gr if it all works it puts a lexicon-fst modified lang directory in `data/recog_langs_gr/MODEL_NAME

Other interesting script:

• common/train_adap_nnet3.sh: The script used to adapt the acoustic model to Egyptian

For myself, I keep the following guidelines for a varikn model:

• For first pass, it should have between 4M and 8M n-grams, for second pass it can be as big as you can train the model. Examples of sizes for different varikn parameters:

  61097228 lm/mgb2/morfessor_f2_a0.01_tokens/varik_aff/arpa-mgb2-0.0005 <- rescore model
  45614982 lm/mgb2/morfessor_f2_a0.01_tokens/varik_aff/arpa-mgb2-0.005
  36084520 lm/mgb2/morfessor_f2_a0.01_tokens/varik_aff/arpa-mgb2-0.01
  22993918 lm/mgb2/morfessor_f2_a0.01_tokens/varik_aff/arpa-mgb2-0.02
   7291900 lm/mgb2/morfessor_f2_a0.01_tokens/varik_aff/arpa-mgb2-0.05 <- Ideal for small
   3422667 lm/mgb2/morfessor_f2_a0.01_tokens/varik_aff/arpa-mgb2-0.08
   2401364 lm/mgb2/morfessor_f2_a0.01_tokens/varik_aff/arpa-mgb2-0.1