I tried to build own language model from own vocabulary.txt and use it for inference from acoustic model v0.5.0
Here is what I did:
lm.binary generation using kenlm(so far nothing to do with deepspeech)
to generate trie (and not to have to compile everything) I used taskcluster.py
See:
https://github.com/mozilla/DeepSpeech/blob/master/USING.rst#using-the-command-line-client which says: python util/taskcluster.py --target .
I had checked out v0.5.0 branch.
then:
./generate_trie alphabet.txt /path-to-own/lm.binary ./output-path/trie
then inference with deepspeech (pip install deepspeech) will result in error: trie mismatch, but continue with inference and so the generated lm.binary will not be used
But, if you use the generate_trie file packed in native_client.tar.gz, like you have said so in this thread, everything will work fine.
So, I think either I’m missing something here, or README.md needs to be updated or taskcluster.py is to blame
Also see: Tune MoziilaDeepSpeech to recognize specific sentences
opinion?