search

srvk / eesen

The official repository of the Eesen project
http://arxiv.org/abs/1507.08240
Apache License 2.0
822 stars 342 forks source link

does eesen has plan for SMBR? #82

Closed xfwu closed 7 years ago

xfwu commented 8 years ago

Hi Yajie

Just wondering if Eesen has this implementation plan? Thank you very much

Best

yajiemiao commented 8 years ago

Hi, discriminative training of CTC models is on our agenda. But we don't have an exact time when it can be delivered.

xfwu commented 8 years ago

Thank you very much!

I have a few questions regarding MPE CTC.

1> Traditional MPE/sMBR employ posterior probabilities like in the SMBR paper

Sequence-discriminative training of deep neural networks

Equation 8 But CTC employs Maximum Likelihood Training. Can I use p(l|x) in

Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks

Equation 15 to replace the P(O|S) in SMBR paper Equation 8?

2>P(W) (in SMBR paper equation 15) supposed to be the language model of the training data, right? What if the training data and the real test data are in different domain, I still need to use the language model of the training data?

Thank you very much

xfwu commented 7 years ago

Hi Yajie

Sorry to bother you with another question:

The basic steps of discriminative training is

  1. get the denlattice
  2. get the alignment
  3. do the discriminative training

For step 1. my understanding is the CTC lattice generated by latgen-faster For step 2. I don't have very clear clue how to get the aligment

Could you please give me some guidance?

Best