Extending Gentle Aligner

Week 10

Customized German Language Model

Using german tuda-de Kaldi recipe

• create lexicon.txt
• phoneGeneration.py Go inside kaldi/egs/gentle/german/ python3 ~/gentle/gentle/phoneGeneration.py proto_dir

UPDATES [POST-GSOC]

• built lm.arpa successfully for lm.arpa — for the sentence:

  ~/kaldi/tools/srilm/lm/bin/i686-m64/ngram-count -text corpus.txt -order 3 -limit-vocab -vocab data/dict/lexicon.txt -lm lm.arpa

• generated langauge files using utils/prep_langauge.sh successfully

  ./utils/prepare_lang.sh --phone-symbol-table de_400k_nnet3chain_tdnn1f_2048_sp_bi/phones.txt german/data/dict '<UNK>' german/data/lang_test/ german/data/lang

• Building grammar

  gzip lm.arpa

  ./utils/format_lm.sh german/data/lang german/lm.arpa.gz german/data/dict/lexicon.txt german/G.fst

• Compiling HCLG.fst

  ./utils/mkgraph.sh --self-loop-scale 1.0 german/G.fst german/model german/

• Decoding using newly compiled HCLG.fst & existing final.mdl [TO DO]

• Getting time-alignments, converting them into visualization format.[TO DO]

• Next: Putting everything together in a Singularity Container Week 11
• Prev: Exploring German Language Models Week 9
• main page

Tools: Kaldi, Python, C, Bash Scripting

Link to GSoC Project Repository