Addons: (Speech recognition) hope to get inside sMBR, MWER tensorflow code

[liangzhenlin123]

System information

• TensorFlow version (you are using): 1.12
• TensorFlow Addons version: Not used yet
• Is it in the tf.contrib (if so, where):No
• Are you willing to contribute it (yes/no):Yes
• Are you willing to maintain it going forward? (yes/no):yes

Describe the feature and the current behavior/state.
Smbr and mwer are important loss functions for speech recognition, and google has published many papers based on these two losses, I hope to get open source code.

Will this change the current api? How?

Who will benefit with this feature?
All speech recognition researchers
Any Other info.

[vijaysumaravi]

+1

[facaiy]

@liangzhenlin123 Hi, would you mind listing the original papers? and if possible, its open source reference implementation in other framework (say, pytorch)

[utkarshsimha]

Would love to work on this - given more information!
Here are some resuources I found for them:
sMBR
MWER

[facaiy]

Welcome, Utkarsh! Feel free to reach out us if you need any help :-)

[prakharcode]

Hi, @utkarshsimha are you taking this? I would like to pick it up if this is open. @facaiy

[facaiy]

Welcome, Prakhar!

@utkarshsimha Hi, Utkarsh, could you please keep us updated? Thanks 😄

[utkarshsimha]

@facaiy @prakharcode - Sorry for the delayed reply.
I have been reading the paper and understanding the mechanics.

Prakhar, if you would like to collaborate, I would be open to that! :)

Thanks!

[liangzhenlin123]

@facaiy @utkarshsimha @prakharcode sorry for the delayed reply :( I have received inspiration elsewhere, and I will finished the N-best MWER in two days, maybe include the sample MWER though it is more difficult than N-best....

[TeaPoly]

@facaiy @utkarshsimha @prakharcode sorry for the delayed reply :( I have received inspiration elsewhere, and I will finished the N-best MWER in two days, maybe include the sample MWER though it is more difficult than N-best....

Can you share MWER for us, or tell us where are your inspiration.

[TeaPoly]

System information

• TensorFlow version (you are using): 1.12
• TensorFlow Addons version: Not used yet
• Is it in the tf.contrib (if so, where):No
• Are you willing to contribute it (yes/no):Yes
• Are you willing to maintain it going forward? (yes/no):yes

Describe the feature and the current behavior/state.
Smbr and mwer are important loss functions for speech recognition, and google has published many papers based on these two losses, I hope to get open source code.

Will this change the current api? How?

Who will benefit with this feature?
All speech recognition researchers
Any Other info.

There is a repo which may be helpful to you: https://github.com/datemoon/tf-code-acoustics. It has some SDT methods , like MPE\MMI.