Building biphone model with predefined phones.txt

[seiten kaku]

Hi,

In steps/train_deltas.sh, there is a context_opts option deciding context width and central position.

What if I modify phones.txt into a biphone version (e.g. A+A 1, A+B 2 ...) and also the related files (nonsilence_phones.txt, text, etc.), and then initiate a model with steps/train_mono.sh, and doing state tying with steps/train_deltas.sh by setting context-width=1 and central-position=0 to train a model, is it still a biphone model?

I have tried it and the WER are likely the same with the one from a normal biphone model with <1% difference. Also if giving lower tying state (higher leaves) and more training data the WER could be lower. 

I come up with this idea while playing with the scripts, but I am not sure if it's a reasonable way to build a biphone (or other CD model) simply because the result looks good. 

Is there any way to verify it? 

Any suggestion will be appreciated, thanks!

[seiten kaku]

Sorry, typo: text -> lexicon.txt

seiten kaku於 2017年8月1日星期二 UTC+8上午2時02分11秒寫道：

[Daniel Povey]

I suppose what you are doing is like a biphone model without tying. A
lot of different things like this were tried a long time ago, when HTK
was popular (it makes all those choices very easy because people were
playing with that stuff a lot at the time it was built). If I were
you I'd use Kaldi's defaults for all that type of stuff.

> --
> 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+...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.