SphinxTrain creates seperate logical models for word-beginning, word-ending, single-word and word-internal triphones. Is there any easy way to get rid of such distinction, i.e. create only one universal model for each triphone, regardless of its position?
Thanks.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
-
2005-04-27
I've never seen any reference to doing this. It seems to be very well integrated into many of the SphinxTrain programs.
But why would you want to? Assuming you wish to use the acouistic models for Sphinx-2/3/4, those recognizers seem to require such triphone models.
cheers,
jerry
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
The reason I want to do this is that I have to use sub-word units (morphemes) as basic units in the language model. While training, on the other hand, I use whole words. Thus, it seems to be wrong to use word-beginning triphone at the beginning of each morpheme when decoding.
Of course, there are other alternatives, such as modifying the lex tree construction, but getting rid of such separation seems to be one of the easiest thing to try out.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
-
2005-04-27
I understand your point. As I understand it, the recognizers will expect to see b-, e-, s-, and internal triphones in the acoustic model, so you would need to disable this in both Sphinxtrain and the recognizer.
Have you considered training and decoding with context-independent models only? There are no such complications with them.
cheers,
jerry
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello,
SphinxTrain creates seperate logical models for word-beginning, word-ending, single-word and word-internal triphones. Is there any easy way to get rid of such distinction, i.e. create only one universal model for each triphone, regardless of its position?
Thanks.
I've never seen any reference to doing this. It seems to be very well integrated into many of the SphinxTrain programs.
But why would you want to? Assuming you wish to use the acouistic models for Sphinx-2/3/4, those recognizers seem to require such triphone models.
cheers,
jerry
The reason I want to do this is that I have to use sub-word units (morphemes) as basic units in the language model. While training, on the other hand, I use whole words. Thus, it seems to be wrong to use word-beginning triphone at the beginning of each morpheme when decoding.
Of course, there are other alternatives, such as modifying the lex tree construction, but getting rid of such separation seems to be one of the easiest thing to try out.
I understand your point. As I understand it, the recognizers will expect to see b-, e-, s-, and internal triphones in the acoustic model, so you would need to disable this in both Sphinxtrain and the recognizer.
Have you considered training and decoding with context-independent models only? There are no such complications with them.
cheers,
jerry