Domain specific Language Model

[bsshruthi22]

We are building an ASR using opensource Hindi/English data using k2-fsa/icefall librispeech recipe.We need to build a Language Model for our domain specific data.Please let us know how to go about this. Thank you

[bsshruthi22]

We are using pruneless stateless transducer7 recipe.

[danpovey]

Hi Sruthi, it was nice to meet you at ICASSP!

I have discussed this with the guys, we are doing 2 things about it:

• One of them (Xiaoyu, I think) will create some howto or documentation about how to do shallow fusion with LODR (which means: divide the LM by a bigram LM estimated on the training transcripts).
• For cases where you have a list of biasing words/phrases: Kangwei/@pkufool will at some point soon make documentation for his implementation of Aho-Corasick and how to use it.

In future, we will have to investigate whether a CTC system interacts better with this kind of LM biasing than RNN-T systems. The guys have already been doing experiments with adding an auxiliary CTC head to the RNN-T system. The RNN-T helps the CTC head learn better (but not vice versa), and I think the CTC WER is nearly as good as the RNN-T one.

[marcoyang1998]

Hi @bsshruthi22 , I'm writing documentation for decoding with language models which should be available very soon. I will update here once I made the PR.

[bsshruthi22]

Thank you very much Dan and @marcoyang1998 .

[bsshruthi22]

Any updates on this @danpovey @marcoyang1998 ?

[pkufool]

@bsshruthi22 https://k2-fsa.github.io/icefall/decoding-with-langugage-models/index.html