It seems it is not possible to create tee models (one state that can optionally be skipped) for short pauses between words using SphinxTrain -- is this correct?
Have any of you found a workaround for this "problem" (not sure if it is a problem)?
Sincerely
Morten & Morten
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Well, usage of T-model as short-pause is probably a unique way of working on speech based on HTK manual. In practice, it is pretty hard to say whether it is the best or not. That is perhaps why, 3.X, 4 and SphinxTrain don't really have support on it.
In practice, what we find is that doing force alignment with selection of data usually gives the best results. Of course, this conclusion is also quite subjected to debate.
In your case, experimenting is usually the best strategy. I found that boostrapping using force-alignment works for both in HTK and Sphinx framework.
Arthur
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi all,
It seems it is not possible to create tee models (one state that can optionally be skipped) for short pauses between words using SphinxTrain -- is this correct?
Have any of you found a workaround for this "problem" (not sure if it is a problem)?
Sincerely
Morten & Morten
Sounds like HTK users to me.
Well, usage of T-model as short-pause is probably a unique way of working on speech based on HTK manual. In practice, it is pretty hard to say whether it is the best or not. That is perhaps why, 3.X, 4 and SphinxTrain don't really have support on it.
In practice, what we find is that doing force alignment with selection of data usually gives the best results. Of course, this conclusion is also quite subjected to debate.
In your case, experimenting is usually the best strategy. I found that boostrapping using force-alignment works for both in HTK and Sphinx framework.
Arthur
Thanks for the quick reply! We'll experiment in order to find the best setup for our system.
Sincerely
Morten & Morten