use of external LM to improve transcriptions

[iggygeek]

Could you explain how to used an external LM, N-gram or RNN to improve the trnascriptions ?

[marcoyang1998]

We provide several decoding methods with external language models. Please have a look at https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md#pruned_transducer_stateless7-zipformer.

[iggygeek]

OK, but why do you use only a BPE-based N-gram LM instead of a word-level N-gram with a lexicon.txt of the type word -> BPE tokens ?

[marcoyang1998]

We support word level N-gram as well, please have a look at here

icefall/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py

Line 101 in 9ddd811

def fast_beam_search_nbest_LG(

[iggygeek]

Well I tried that a word-level ARPA LM with a lexicon.txt of the type word -> BPE tokens using the BPE models used for training.
I validated using validate_lang_bpe.py, but I got empty hypotheses when decoding with fast_beam_search_nbest_LG ...

[marcoyang1998]

Could you please share the decoding command you are using?

BTW, using neural network LMs usually gives a larger performance improvement. If you want to have optimal performance, we recommend you use modified_beam_search_LODR.

[iggygeek]

I think the problem was that the script generating LG was using a default path for the ARPA LM, whereas I thought that it would looked for it in the lang_bpe_XXX directory...