Shallow fusion of LSTM LM

[rwth-i6]

Shallow fusion of LSTM LM

[albertz]

@kazuki-irie Can you upload your shallow fusion config also (here)?

@akshatdewan Yes, sorry, we forgot to upload that config. You will find a similar config for shallow fusion for Switchboard here. It should not be hard to adapt this in the same way for LibriSpeech.

[akshatdewan]

Thanks @albertz I will try modifying the SWB config file.

[kazuki-irie]

I have added the config.

[albertz]

Thanks! I'm closing this now. Feel free to reopen if there are further issues.

[akshatdewan]

Thank you! I just have a question regarding the input to the lm subnetwork -
I am having troubles understanding what it means when it says "from":["prev:output"] here. Could you please help me understand where the input to the layer lm_out is coming from?
Thanks

[albertz]

It means that the input to the LM is the output of the previous frame.

[akshatdewan]

Thanks! And I am sorry, I missed it here
In the subnetwork, you can access outputs from layers from the previous time step when they are referred to with the "prev:" prefix.

[yanghongjiazheng]

@akshatdewan
Hi,

I wang to train an LSTM LM using the config. But I have no idea about the exactly form of the /work/asr3/irie/data/librispeech/lm_bpe/librispeech-lm-norm.bpe.txt.gz . So I can not shape my data. And also I coming to the trouble of def cf(filename) , err_msg = "No such file or directory: 'cf'".
Can you please help me out.
Thanks for your help.

[albertz]

@yanghongjiazheng See my answer in #21, and please let's keep any further discussion/question in there.