About AMI experiments - using IHM alignments for SDM/MDM?

[Searcher Ray]

Dear all,

I'm currently playing with the running scripts in egs/ami/s5(b). However, I found a thing that is pretty confusing to me. At stage 11/12 (11 is default running setting as chain-based TDNN has been proved to be better than nnet3), there is such command (bold one):

===

ali_opt=

  [ "$mic" != "ihm" ] && ali_opt="--use-ihm-ali true"

  local/chain/run_tdnn.sh $ali_opt --mic $mic

===

So if I'm not wrong, this command says something like `use state alignment generated by IHM if the microphone option is not IHM'. This is quite confusing as we're trying to use alignments from another different case for NN training - again, at least to me, maybe not at all for others.

Am I understand this wrong or it is what it is? I can't find the solution directly on relevant papers as well. Anybody can give any sensible explanations on this?

Thanks for helping!

[Daniel Povey]

If you check my publications page there will be a paper with Vijay as first author that mentions this.

The individual headset microphones give better alignments of the transcripts.  The issue we're solving, is that due to other speakers being audible and there being a lot of noise in the other microphone conditions, the alignments can have gross errors-- words in completely the wrong place.  If we use the alignments from the IHM condition to train for those setups, we get better systems.

Dan

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
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+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.