Re: [kaldi-help] End-to-end speech recognition using lattice-free MMI

[Daniel Povey]

Do

find . -name 'run_e2e*.sh'

from the kaldi directory

On Mon, Aug 8, 2022 at 11:37 PM Sage Khan <class...@gmail.com> wrote:

I was going through 2018 Interspeech publication "End-to-end speech recognition using lattice-free MMI" (Hossein Hadian1;2 , Hossein Sameti1, Daniel Povey2;3, Sanjeev Khudanpur2;3)

I am unable to find any script or recipe for this model. I wanted to use it on my dataset.

Can anyone please help out?

--
Go to http://kaldi-asr.org/forums.html to find out how to join the kaldi-help group
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/9422ef31-8a75-4a8c-9f15-ced61cc7966fn%40googlegroups.com.

[Sage Khan (Sage Khan)]

I was going through KALDI DOCS on Chain Training as well as through this blog On lattice free MMI and Chain models in Kaldi (desh2608.github.io).

I wanted to know if this diagram summarizes whole LFMMI Training in the Diagram above. Can you help point me out where exactly the whole CNN-TDNN pipeline comes in where 40x6 speech features of 1500ms segments with 200 ivectors and 40 Mel Filter bank features are taken as input, convoluted etc?

This is the actual part I am struggling to understand.

[Jan Yenda Trmal]

hi, the Mel Feats and TDNN/CNN is essentially the input and structure of the network. The LF-MMI is the training objective (i.e. how the weights are set during the training). So everything fits within the box "Neural network training"

y.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/4a9179a1-136a-455f-be98-a4d9b0e8af62n%40googlegroups.com.

[Sage Khan (Sage Khan)]

Thank you so much :)