Minmum Word Error Rate Training（MWER）

[ColainCYY]

Hi! Is there a guide to implement MWER Training in lingvo? I have tried to write the code from scratch, but it is a really difficult to me. I'm trying the beam-srarch based N-best MWER Training as a finetune for a trained MLE model.

[drpngx]

@rprabhavalkar can't remember where we are on the open-sourcing of MWER?

On Thu, Jul 4, 2019 at 12:28 PM ColainCYY notifications@github.com wrote:

Hi! Is there a guide to implement MWER Training in lingvo? I have tried to write the code from scratch, but it is a really difficult to me. I'm trying the beam-srarch based N-best MWER Training as a finetune for a trained MLE model.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tensorflow/lingvo/issues/118?email_source=notifications&email_token=AE75E3NRKXZFFUIGOYATIWDP5V35XA5CNFSM4H5SCEC2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4G5I5H3Q, or mute the thread https://github.com/notifications/unsubscribe-auth/AE75E3M6V2JKOQLE7ISUQ7TP5V35XANCNFSM4H5SCECQ .

[Some-random]

I'm doing something similar and an open source implementation of MWER will be really helpful for us!

[rprabhavalkar]

Thanks for the question. We are not actively working on open-sourcing the MWER implementation at the moment. That being said, I can look into this a bit more to see if this is something that we can do relatively quickly. I'll update this thread once I have more up to date information.

[zh794390558]

Watch this.

[qgzang]

mark

[chenjiasheng]

@ColainCYY @drpngx @Some-random @rprabhavalkar @zh794390558 @qgzang I have implemented mWER loss here: https://github.com/chenjiasheng/mwer.

[qgzang]

@ColainCYY @drpngx @Some-random @rprabhavalkar @zh794390558 @qgzang I have implemented mWER loss here: https://github.com/chenjiasheng/mwer.

thx