Phonemes recognition

[Roman]

Hello!

To see how well the acoustic model works, I decided to see phonemes recognition on test set.

The closest to the solution of my problem was the use of the utility "post-to-phone-post".

But this utility is dependent on the language model.

Is there a way to get phonemes recognition without affecting the language model?

Thanks!

[Daniel Povey]

It doesn't really use the model, it only reads it in order to have available the mapping from transition-ids to phones.

Dan

--
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.

[Roman]

Sorry, I incorrectly described the experiment.

To get the "post" I am using "lattice-to-post", and then "post-to-phone-post".

A utility "lattice-to-post" depends on language model.

Can I get a "post" without affecting language model?

Thanks!

понедельник, 9 ноября 2015 г., 21:34:35 UTC+3 пользователь Dan Povey написал:

[Daniel Povey]

lattice-to-post does not read a language model (which is clear because it doesn't have a command line argument for one).  It accepts a --lm-scale option because it needs to know how to scale the language model versus acoustic model weights in the lattice.  This will be whatever phone-level language model you decoded with.  There is always a language model, even if it's a very simple one.  For instance a phone loop or a bigram phone language model.

Dan

[tushev...@gmail.com]

Hello.
I have same problem. If I use different LM weights, I get different results. Is it correct to set lmwt=0? Or I should set lmwt with best WER?

[Daniel Povey]

In most circumstances, lmwt=0.1 is appropriate.  Or you could set it to the inverse of the best acoustic scale, like 1/11 or whatever.

Dan

On Wed, Feb 24, 2016 at 5:34 AM, <tushev...@gmail.com> wrote:

Hello.
I have same problem. If I use different LM weights, I get different results. Is it correct to set lmwt=0? Or I should set lmwt with best WER?

--
You received this message because you are subscribed to the Google Groups "kaldi-help" group.

[tushev...@gmail.com]

Thank you for reply!

[Сергей Тушев]

Hello

I've got strange decode result. If I use small LM (nwords 38k), I get one phoneme string, and when I use big LM with nwords 480k, I get another phoneme string. But AM model is same. Is this ok? I thought, that phoneme string should not depend on LM.

Thank you

[Daniel Povey]

It definitely depends on the LM. Read the HTK book or search for
"Gales Young HMM" to understand how HMM-based speech recognition
works.
Dan

[Сергей Тушев]

I try to explain my problem.

I need to test my AM, check it's quality. How could I do it? 

I suppose, that if I get phoneme string (instead of words), I could get phone accuracy and correct. 

If I use very simple unigram LM and 3gram LM, phoneme string will be different? How can I avoid it?

In other words, is it possible in Kaldi to get phone error rate (PER)?

That's the way I get phoneme string:

lattice-to-post --acoustic-scale=0.1 --lm-scale=0.1 ark:"gunzip -c $lattice|" ark:- | post-to-phone-post $model ark:- ark,t:$lattice.post.int

Is it write?

[Daniel Povey]

People often ask this, but it's not really possible [or at any rate,
it doesn't make sense] to check an AM in the absence of an LM.
Dan