--ivector-silence-weighting in NNET3 ASR model training with iVectors

[Filip Jurcicek]

Hi all, 

I noticed that there is are options  --ivector-silence-weighting.* for nnet3 online decoders. 

I am wondering if there any options in nnet3 model training like that so that the model is trained with weighted ivectors and preventing a mismatch between training and decoding setups. I have not found anything like that. 

I am using steps/online/nnet2/extract_ivectors_online.sh to extract ivectors for ASR training and it does not seem to use any information about speech / non-speech information (e.g. from alignments).

If it is not implemented yet, do you think that it would be helpful? I assume that it may have an impact when the training data has a lot of silence/noise segments.

Thanks!

Best regards,

Filip

[Daniel Povey]

Personally I doubt that it would be that helpful to do it in training.. the intuition is that if it sees both the speech and silence during training it will be more robust to any silence that *does* creep in at test time.  Plus, knowledge about the silence properties is important for ASR as well as knowledge about the speech properties.  You could try and see- but it might be tricky to integrate into the training pipeline.

The silence-weighting thing is only useful or important for domain-mismatched conditions (like the ASPIRE challenge).   I'm not aware that it's helpful for same-domain testing.

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.

[Filip Jurcicek]

I see. Thank you for your insight. 

All the best, 

Filip

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.