Large scale language model
Caution! This tutorial is out of date. We recommend using KenLM as it is free software unlike SRILM. It is also faster.
Building a large scale language model for domain-specific transcription
Language model describes the probabilities of the sequences of words in the text and is required for speech recognition. Generic models are very large (several gigabytes and thus impractical). Most recognition systems have models tuned to the specific domain. For example, medical language model describes medical dictation. If you are looking for your domain you most likely will have to build the language model for that domain yourself. You can mix that specific domain with generic domain to get some fallback, but specific domain is still needed. Generic language models are created from large texts.
The language modeling toolkit is http://www.speech.sri.com/projects/srilm, it contains most tools for language modeling. In modern systems there is a lot of success with neural-network based RNNLM language models. Those are supported in rnnlm toolkit.
The first step of language model building is a collection of the data. The amount of data you need depends on the domain and vocabulary. Usually for a good model you need a significant amount of texts - at least 100mb. You can get this text by transcribing existing recordings, collecting data from the web, generating it artificially with scripts. The most valuable data is a real-life data anyway.
Once data is collected it must be cleaned - punctuation removed, sentences split on lines, numbers expanded to text representation. This can be done with scripting language
In machine learning it is practical to control the quality of the model you build. For that you need to separate test set from training set and use test set for evaluation. The practical split is 1 to 10.
The metric to use for evaluation is called perplexity. It controls the quality of the language model. The smaller perplexity is, the smaller is WER. You can calculate perplexity with SRILM:
ngram -lm your.lm -ppl test.txt
Most other toolkits can also calculate perplexity.
Another parameter is OOV rate. SRILM also prints that with -ppl option. The OOV rate should not be large, the WER because of OOV is usually double the OOV rate. For example if you have 3% OOV rate you add extra 6% to WER. So you need to try to minimize OOV rate by extending your training set.
Initial model estimation
Language model can be estimated with different parameters. The parameters depend on a task and require some understanding for underlying algorithms. For example, for book-like texts you need to use Knesser-Ney discounting. For command-like texts you should use Witten-Bell discounting or Absolute discounting. You can try different methods and see which gives better perplexity on a test set.
The command is
ngram-count -kndiscount -text text.txt -lm text.lm
Language model interpolation with generic model
To improve language model coverage you can mix LM with generic LM created from WEB texts. There are few of them:
Language model from Cantab Research trained on gigaword corpus
To perform the mix you need to estimate mixture parameters. For that you can evaluate model on test set:
ngram -lm your.lm -ppl test.txt -debug 2 > your.ppl ngram -lm generic.lm -ppl test.txt -debug 2 > generic.ppl compute-best-mix your.ppl generic.ppl ngram -lm your.lm -mix-lm generic.lm -lambda `<factor from above>` -write-lm mixed.lm
The mixed LM usually has slightly better perplexity than specific LM.
Language model pruning
Most language models are large and it is impractical to use them in decoder. For example, you can not pack 1Gb LM into WFST decoder. For that reason you can prune them to reduce their size:
ngram -lm mixed.lm -prune 1e-8 -write-lm mixed_pruned.lm
You can try different factors to get right model size. Usually model must be within 30mb. You perform decoding with pruned model instead of full model.
To compensate reduction of accuracy due to pruning you still can use original model in lattice rescoring mode to get the best accuracy. Most decoders can dump lattices and lattices can be rescoring with very large models.
You can also use RNNLM in rescoring model to rescore n-best lists to get the best accuracy.
You might need to extend the dictionary with the words from the language model because they might be missing. You can use G2P tool like phonetisaurus for that, however, remember that such tools are usually inaccurate. For that reason it is recommended to manually review G2P output and correct it if needed.