I’ve worked on DS 0.4.1 and the voice model is working nicely, as suggested by @tuttlebr(on unfreezing).
Now, I’m planning on a language model interpolation; i.e. something like:
LM(final) = W1*Generic LM + W2*Domain Specific Language Model
(Where W1 & W2 are respective weights of the language model)
Is it possible to do something like this on deepspeech, and if so, how do I train the system to select W1 & W2 (or has it to be institution based?)
Additionally, it seems it’s easier to get the final LM using SRILM. Now, will it be ok, if we generate an arpa in SRILM, convert it to binary and then use deepspeech util to convert it to trie?
Finally, going through the code, I understand there’s a hyperparameter lm_alpha. Could you explain what that does?
And can you possibly guide me to the codes which are used while decoding and/or the interface of the final voice model with the LM model?