Decoding using Domain Language Model in sherpa

[bsshruthi22]

We have trained a zipformer model with 3500 hours .We built a domain specific language model using n gram model using build_lm.py script. We found the performance same with and without LM. We tried decoding with sherpa using LG.pt 1.Please advice regarding performance. 2.How to use RNN language model in sherpa?

Thanks in advance.

[bsshruthi22]

@csukuangfj @danpovey Could you kindly give some info about this?

[pkufool]

@bsshruthi22 As for the performance, what's your lm score? and did you see some patterns of the decoding results with LM. It is hard to give you the reason directly, you have to debug step by step by yourself, for example, first confirm whether an LM was used during the decoding.

[pkufool]

How to use RNN language model in sherpa?

There is a fast_beam_search_decoder in sherpa which accepts an LG as decoding graph.