search

wenet-e2e / wenet

Production First and Production Ready End-to-End Speech Recognition Toolkit
https://wenet-e2e.github.io/wenet/
Apache License 2.0
4.14k stars 1.07k forks source link

gigaspeech LM notworking with gigaspeech U2PP model #1695

Closed jayakrishnanmm closed 1 year ago

jayakrishnanmm commented 1 year ago

I am using pre-trained/gigaspeech/20210728_u2pp_conformer_libtorch as acoustic model and want to use gigaspeech LM from https://huggingface.co/wgb14/gigaspeech_lm while decoding. I built TLG.fst using the instructions from https://github.com/wenet-e2e/wenet/blob/main/docs/tutorial_librispeech.md (Stage 7) I have used the lexicon from the same source and using a bpe model to split it.

./build/bin/decoder_main \ --chunk_size -1 \ --wav_path $wav_path \ --model_path $model \ --unit_path $units \ --rescoring_weight $rescoring_weight \ --ctc_weight $ctc_weight \ --reverse_weight $reverse_weight \ $wfst_decode_opts \ --result $result/youtube.txt &> $result/out.log

Howver, I am getting a totally wrong junk text, onlly few words are meaningful. When I decode the model without LM, it is ok. What is happening here ?

robin1001 commented 1 year ago

Did you get the right output when LM is not used?

jayakrishnanmm commented 1 year ago

Yes.

robin1001 commented 1 year ago

There must be some error when you building the FST. please make sure the modeling units of gigaspeech model(the model you download) is used for building the FST.

xingchensong commented 1 year ago

This may help you https://github.com/wenet-e2e/wenet/issues/1673

jayakrishnanmm commented 1 year ago

Thanks. Now working

naomibb commented 1 year ago

May I ask how did you work in the end? and must use the bpe model. where can i get the bpe model? thanks

jayakrishnanmm commented 1 year ago

The bpe model is trained using the train text. You refer the run.sh under examples/gigaspeech/s0