xubuild / lm_build

Adapting your own Language Model for Kaldi

Kaldi Language Model Building

Adapting Your Own Language Model

Instructions to learn about building a Kaldi language model based on your own text.

When you clone this code into a Kaldi experiment like …/kaldi-trunk/egs/tedlium/s5 you get a folder lm_build/ with tools and examples of how to adapt and train a language model based on your own training text file.

Adding New Vocabulary Words to the Lexicon

The new script run_adapt.sh helps make LM adaptation much easier now.

• Method 1: manually create a file newwords.txt in the lm_build working folder, into which you place new words (not already in the lexicon in TEDLIUM.152k.dic) Pronunciations will be automatically generated and added to the dictionary.
• Method 2: Automatic candidate OOV words are generated when you run run_adapt.sh in the file candidate_oovs.txt. This candidate list of new words contains all words found in the training text not already in the dictionary (OOV words) that appear more than once. Rename this file newwords.txt and run run_adapt.sh again to use all these words with a frequency greater than 2. Or edit newwords.txt having a look at oov-counts.txt to see the word frequency counts and help you iteratively refine the dictionary
• (optionally) add to the example_txt training text file some examples that use the new words. Hint: you may need to repeat these LM adaptation sentences between 50 and 100 times for the transcriber to recognize and produce them as output.
• Run the script run_adapt.sh. This will do several things, but the end result will be a new composed decoding graph TLG.fst in the output folder data/lang_phn_test/
• Point your Eesen Transcriber setup to use the resulting graph, for example by setting this value in /vagrant/Makefile.options

GRAPH_DIR?=$(EESEN_ROOT)/asr_egs/tedlium/v2-30ms/lm_build/data/lang_phn_test

Adding your own pronunciations

This process makes use of the CMU Lexicon Tool to generate dictionary entries with phonetic pronunciations for unseen words. These may not always be correct. An alternative approach (Method 3?) Add your own words and pronunciations directly to TEDLIUM.152k.dic first - perhaps pattern matching parts of pronunciations from similar words. It is also possible to have more than one pronunciation, e.g:

zydeco Z AY D EH K OW
zydeco(2) Z IH D AH K OW
zydeco(3) Z AY D AH K OW

FAQ

How some of the scripts work
Deterministic (tiny) LM Building
Adding Technical Words to Dictionary
More Details About LM Building