I want to conduct an experiment wherein I want to ensure that certain words(from the reference) are also present in the hypothesis(after decoding) at their corresponding time stamps. We can assume that the word and its time stamp are available in advance.
To do this, I was wondering if I could boost the scores of such words in the lattice, in such a manner that if you re score the lattice, those words are always found in the hypothesis.
I peeked into the lattice and couldn't come up with how to pin down the words. Would you have any leads as to how to proceed with the experiment?
Thanks,
Yash
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
This is difficult because those words might not even be in the lattice
in the first place, and even if they are, you'd have to define exactly
what you mean by "at the same time".
The program lattice-boost-ali will modify the costs of transition-ids
that have the same phone as the reference; if you have a reference
alignment, you could give it a negative boosting factor (--b=-0.5 for
instance) and it will do something vaguely related to what you said.
Dan
I want to conduct an experiment wherein I want to ensure that certain
words(from the reference) are also present in the hypothesis(after decoding)
at their corresponding time stamps. We can assume that the word and its time
stamp are available in advance.
To do this, I was wondering if I could boost the scores of such words in the
lattice, in such a manner that if you re score the lattice, those words are
always found in the hypothesis.
I peeked into the lattice and couldn't come up with how to pin down the
words. Would you have any leads as to how to proceed with the experiment?
Hi Dan et. al.,
I want to conduct an experiment wherein I want to ensure that certain words(from the reference) are also present in the hypothesis(after decoding) at their corresponding time stamps. We can assume that the word and its time stamp are available in advance.
To do this, I was wondering if I could boost the scores of such words in the lattice, in such a manner that if you re score the lattice, those words are always found in the hypothesis.
I peeked into the lattice and couldn't come up with how to pin down the words. Would you have any leads as to how to proceed with the experiment?
Thanks,
Yash
This is difficult because those words might not even be in the lattice
in the first place, and even if they are, you'd have to define exactly
what you mean by "at the same time".
The program lattice-boost-ali will modify the costs of transition-ids
that have the same phone as the reference; if you have a reference
alignment, you could give it a negative boosting factor (--b=-0.5 for
instance) and it will do something vaguely related to what you said.
Dan
On Wed, Dec 31, 2014 at 2:55 PM, Yash y91@users.sf.net wrote:
Hey Dan,
Is it possible to insert a word in the kaldi lattice, given that I already know that it exists at a certain time?
Thanks!
Last edit: y91 2014-12-31
No.
On Wed, Dec 31, 2014 at 3:52 PM, Yash y91@users.sf.net wrote: