lattice - phones sequence

[emelie....@gmail.com]

Hi! 

If I have (sort of) followed the standard rm recipe (but with different training data), is it possible to after the monophonic training and decoding see a lattice for the phone sequence in an utterance? 

Also I have by the show_lattice generated word lattices, but I can't seem to see any probabilities in the graph, is that something that can be displayed also?  

Thanks a lot, 

Emelie 

[Daniel Povey]

If I have (sort of) followed the standard rm recipe (but with different training data), is it possible to after the monophonic training and decoding see a lattice for the phone sequence in an utterance? 

show-alignments lets you view the alignments in a human-readable way. 

Also I have by the show_lattice generated word lattices, but I can't seem to see any probabilities in the graph, is that something that can be displayed also?  

it's not the probabilities as such that are in the lattice, the probabilities can be worked out though.   see lattice-arc-post, you can get more readable output from that, but take care of probability scales.   read documentation on lattices at kaldi.asr.org- I am busy.
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.

[Emelie Kullmann]

Yes thanks, I’ve run that too, that was very helpful! But this is for the decoding, I should reformulate, is there a way to see the phonetic sequence for an utterance in the same way as you can see the word sequence for an utterance in show_lattice? 

Something like this 

or do you think it would be possible to extract that from the model? I’m not a very skilled programmer but if show_lattice could be a start I could probably go from there, if you think it would be possible to extract!

Thanks again, I will definitely do that! 

Emelie 

You received this message because you are subscribed to a topic in the Google Groups "kaldi-help" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/kaldi-help/ve75M85R_-8/unsubscribe.
To unsubscribe from this group and all its topics, send an email to kaldi-help+...@googlegroups.com.

[Daniel Povey]

lattice-to-phone-lattice will modify the lattice to have phones as the labels.

lattice-1best and lattice-prune can be used to keep just the 1-best path or prune away bad-scoring paths.  (be careful with the acoustic scale).

lattice-to-fst can be used to convert to a regular FST which things like fstdraw would work with.

You need to read the documentation as I don't have time to give detailed guidance.

[Emelie Kullmann]

alright, thank you very much dan! I will get right on it!  

On 05 Apr 2016, at 21:20, Daniel Povey <dpo...@gmail.com> wrote:

lattice-to-phone-lattice will modify the lattice to have phones as the labels.

lattice-1best and lattice-prune can be used to keep just the 1-best path or prune away bad-scoring paths.  (be careful with the acoustic scale).

lattice-to-fst can be used to convert to a regular FST which things like fstdraw would work with.

You need to read the documentation as I don't have time to give detailed guidance.

On Tue, Apr 5, 2016 at 1:46 PM, Emelie Kullmann <emelie....@gmail.com> wrote:

Yes thanks, I’ve run that too, that was very helpful! But this is for the decoding, I should reformulate, is there a way to see the phonetic sequence for an utterance in the same way as you can see the word sequence for an utterance in show_lattice? 

Something like this 

<Screen Shot 2016-04-05 at 19.42.28.png>