Multiple pronunciations for words and oov words

[yair shachar]

Hi all, I have two questions regarding the lexicon file that we create when training the acoustic model,

1) Multiple pronunciations for words: 

If I have words that might have several pronunciations, for example:

the z eh

the z iy

the d eh

etc'...

And words that have generated pronunciations via g2p.

In the training data. I do not know which of these pronunciations was said, could this heavily affect the training and is there a more optimal way to deal with multiple pronunciations for words? Is it better to have only one pronunciation per word?

2) Learning oov words - as part of the training the 'spn' phoneme is also learned, for what I understand this is done by choosing words that will not be present in the lexicon when training. is there a smart way to choose these words and/or a good rule of thumb for the percentage of words to choose such that they will be learned as oov words in order for the model to learn spn well?

Many thanks

[Jan Yenda Trmal]

generally, we do not care and we let the alignment phase deal with the issue of finding the proper pronunciation

I think it's generally better (with respect to WER) to not overgenerate the g2p pronunciations, but you might need to run your own evaluation.

2) recognizing OOV properly is still not solved and no method currently in kaldi gives you good-enough performance, IMO. Usually, in kaldi, the OOV model is used only for training, so that you can use sentences with OOV during training.

y.

--
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/fd57aac1-028d-41a9-9cf2-404b38fb3940n%40googlegroups.com.

[yair shachar]

Thanks a lot. Is no "best practice" for usage of OOV words during training? i.e if no words would be used as OOV in training, wouldn't that mean that the 'spn' phoneme wouldn't be learnt?