In the past there was a script to do it, i think, but now I can’t find a way to generate a trie. I need to pass it the Model.enableDecoderWithLM(…) method. Though I did notice for training i had to remove the lm_trie_path argument which i previously used but now isn’t supported.
I generated a language model using the generate_lm.py script which produced a lm.binary file but no corresponding trie, as far as I can see, although there is now a vocab-5000000.txt file in the directory.
Although now i’m looking, the enableDecoderWithLM method doesn’t even exist anymore. Interesting