CMU Sphinx / Forums / Help: sphinx lattice to WCN conversion

shamsa abid - 2011-08-05

Hi
Im working on converting my pocketsphinx lattice into a WCN/sausage.
Ultimately,i want all possible candidates for a single word utterance made
available to choose from.

Once i get the lattice object, what should my next steps be?
I was planning to use SRILM lattice-tool for this, however, it accepts only
PFSG and HTK style lattices.
How do i transform sphinx lattice to a SRILM compatible lattice?

Regards,
Shamsa.

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

Nickolay V. Shmyrev - 2011-08-06

pocketsphinx can output lattices in HTK format

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

shamsa abid - 2011-08-21

I am using pocketsphinx as decoder for my android application.
and I intend to use SRILM for sausage generation.

I was able to generate the htk format lattice using the -outlatfmt htk
parameter in the pocketsphinxbatch command

i later gave that file as input to the srilm lattice-tool (on cygwin prompt)
and got a mesh file

Now i want this mesh available in my android application. In my android app i
have got a lattice object and i am able to access the words on the nodes of
the lattice.

I need to know which srilm file has the method that would take my lattice
object as input and return me a lattice which is in sausage form. (or whether
it is possible or not? because i studied the code and i could see lattice
being manipulated as a file)

Your comments and guidance is greatly appreciated.

Regards

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

Nickolay V. Shmyrev - 2011-08-22

Answer from Andreas Stolke on srilm-users mailing list

You would have to construct the lattice data structure in memory. The best way
to learn how to do that (I admit it is somewhat involved) is to read the code
that reads an HTK lattice file into memory. You can find this in
$SRILM/lattice/src/HTKLattice.cc, function Lattice::readHTK().

A more straightforward, if somewhat less efficient way would be to write the
textual HTK lattice representation to a string object, and then "read" from
that string using Lattice::readHTK(). For this purpose the File object can be
constructed from a string, see the comment in $SRILM/misc/src/File.h . I
believe you need to get the very latest (beta) version of SRILM for this to
work.

Andreas

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

shamsa abid - 2011-09-14

The ps_lattice_write method for pocketsphinx isnt working for me because im
running it for my android app. The file name that i specify is an android

/mnt/sdcard/edu.cmu.pocketsphinx

folder. The problem may be that the C code uses a file object.
The following code is from ps_lattice.c

if ((fp = fopen(filename, "w")) == NULL) {
E_ERROR("Failed to open lattice file '%s' for writing: %s\n", filename,
strerror(errno));
return -1;
}

I always get -1 returned to my calling code.

I desperately need to get the lattice written out so i can see what it looks
like after the decoder outputs it. Meanwhile Im also trying to traverse the
lattice using the following code.

while (nodeiterator.getnode().hyp(this.lattice)!=null )
{
nodeiterator=nodeiterator.next();
if (nodeiterator!=null)
{
n1=nodeiterator.getnode().hyp(this.lattice).getHypstr();
System.out.println(n1);// prints the word at the node

}else
break;
if(n1=="")
break;

int dummy=0;
}

This gives me the following output:

09-14 12:23:50.984: INFO/System.out(319): <sil>
09-14 12:23:50.984: INFO/System.out(319): <sil>
09-14 12:23:51.013: INFO/System.out(319): <sil>
09-14 12:23:51.033: INFO/System.out(319): <sil>
09-14 12:23:51.075: INFO/System.out(319): <sil>
09-14 12:23:51.075: INFO/System.out(319): <sil>
09-14 12:23:51.154: INFO/System.out(319): <sil>
09-14 12:23:51.194: INFO/System.out(319): AN(2)
09-14 12:23:51.235: INFO/System.out(319): <sil>
09-14 12:23:51.235: INFO/System.out(319): <sil>
09-14 12:23:51.253: INFO/System.out(319): <sil>
09-14 12:23:51.264: INFO/System.out(319): AN(2)
09-14 12:23:51.293: INFO/System.out(319): <sil>
09-14 12:23:51.293: INFO/System.out(319): <sil>
09-14 12:23:51.314: INFO/System.out(319): THE
09-14 12:23:51.334: INFO/System.out(319): OUT
09-14 12:23:51.354: INFO/System.out(319): AND(2)
09-14 12:23:51.383: INFO/System.out(319): A
09-14 12:23:51.383: INFO/System.out(319): IT
09-14 12:23:51.424: INFO/System.out(319): <sil>
09-14 12:23:51.433: INFO/System.out(319): AT
09-14 12:23:51.463: INFO/System.out(319): IT
09-14 12:23:51.463: INFO/System.out(319): IT.
09-14 12:23:51.485: INFO/System.out(319): IT?
09-14 12:23:51.485: INFO/System.out(319): <sil>
09-14 12:23:51.514: INFO/System.out(319): A(2)
09-14 12:23:51.514: INFO/System.out(319): IT.
09-14 12:23:51.553: INFO/System.out(319): IT?
09-14 12:23:51.553: INFO/System.out(319): IT
09-14 12:23:51.574: INFO/System.out(319): AN
09-14 12:23:51.574: INFO/System.out(319): IN
09-14 12:23:51.604: INFO/System.out(319): TO(2)
09-14 12:23:51.604: INFO/System.out(319): A(2)
09-14 12:23:51.624: INFO/System.out(319): A(2)
09-14 12:23:51.624: INFO/System.out(319): <sil>
09-14 12:23:51.663: INFO/System.out(319): <sil>
09-14 12:23:51.674: INFO/System.out(319): A(2)
09-14 12:23:51.693: INFO/System.out(319): <sil>
09-14 12:23:51.693: INFO/System.out(319): <sil>
09-14 12:23:51.723: INFO/System.out(319): <sil>
09-14 12:23:51.723: INFO/System.out(319): IN
09-14 12:23:51.743: INFO/System.out(319): TO(2)
09-14 12:23:51.743: INFO/System.out(319): AND(2)
09-14 12:23:51.763: INFO/System.out(319): <sil>
09-14 12:23:51.774: INFO/System.out(319): HAVE
09-14 12:23:51.804: INFO/System.out(319): THAT
09-14 12:23:51.804: INFO/System.out(319): I'VE
09-14 12:23:51.815: INFO/System.out(319): AT
09-14 12:23:56.863: INFO/System.out(319): <sil>
09-14 12:23:56.873: INFO/System.out(319): A
09-14 12:23:56.903: INFO/System.out(319): </sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil>

and my code for traversal using exit nodes looks like
this.lattice= new Lattice(this.ps);
node = new Latnode ();

nodeiterator=new Latnode();
nodeiterator=node.LatIterator(this.lattice);

//my working code for iteration in lattice

//code which iterates on node exits, currently giving repetitions
node=nodeiterator.getnode();
latlinkiter=new LatlinkIterator();

while (node!=null )
{
latlinkiter=node.nodeExits();

while (latlinkiter!=null)
{
String linkword=latlinkiter.LinkIterLink().LinkWord(this.lattice);
latlinkiter=latlinkiter.LinkIterNext();
System.out.println(linkword);

}
nodeiterator=nodeiterator.next();
if(nodeiterator!=null)
node=nodeiterator.getnode();
else
break;

}

and my results look like

09-14 12:23:56.973: INFO/System.out(319): <sil>
09-14 12:23:57.004: INFO/System.out(319): <sil>
09-14 12:23:57.004: INFO/System.out(319): <sil>
09-14 12:23:57.024: INFO/System.out(319): <sil>
09-14 12:23:57.024: INFO/System.out(319): <sil>
09-14 12:23:57.053: INFO/System.out(319): <sil>
09-14 12:23:57.053: INFO/System.out(319): <sil>
09-14 12:23:57.084: INFO/System.out(319): <sil>
09-14 12:23:57.084: INFO/System.out(319): <sil>
09-14 12:23:57.104: INFO/System.out(319): <sil>
09-14 12:23:57.104: INFO/System.out(319): <sil>
09-14 12:23:57.133: INFO/System.out(319): <sil>
09-14 12:23:57.133: INFO/System.out(319): <sil>
09-14 12:23:57.154: INFO/System.out(319): AN(2)
09-14 12:23:57.154: INFO/System.out(319): AN(2)
09-14 12:23:57.174: INFO/System.out(319): AN(2)
09-14 12:23:57.174: INFO/System.out(319): <sil>
09-14 12:23:57.203: INFO/System.out(319): <sil>
09-14 12:23:57.203: INFO/System.out(319): <sil>
09-14 12:23:57.224: INFO/System.out(319): <sil>
09-14 12:23:57.224: INFO/System.out(319): <sil>
09-14 12:23:57.254: INFO/System.out(319): <sil>
09-14 12:23:57.254: INFO/System.out(319): AN(2)
09-14 12:23:57.263: INFO/System.out(319): AN(2)
09-14 12:23:57.273: INFO/System.out(319): AN(2)
09-14 12:23:57.273: INFO/System.out(319): AN(2)
09-14 12:23:57.304: INFO/System.out(319): <sil>
09-14 12:23:57.304: INFO/System.out(319): <sil>
09-14 12:23:57.324: INFO/System.out(319): THE
09-14 12:23:57.324: INFO/System.out(319): THE
09-14 12:23:57.343: INFO/System.out(319): THE
09-14 12:23:57.343: INFO/System.out(319): THE
09-14 12:23:57.353: INFO/System.out(319): THE
09-14 12:23:57.374: INFO/System.out(319): THE
09-14 12:23:57.374: INFO/System.out(319): THE
09-14 12:23:57.403: INFO/System.out(319): THE
09-14 12:24:19.114: INFO/System.out(319):
09-14 12:24:19.114: INFO/System.out(319):
09-14 12:24:19.114: INFO/System.out(319):
09-14 12:24:19.114: INFO/System.out(319):
~~09-14 12:24:19.114: INFO/System.out(319):~~ </sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil></sil>

Well i have cutdown so many lines of the results, because there were just too
many lines.

Now whats really bugging me is
1) I am unable to write the complete lattice file on android, i even tried
giving my hard disk path, that didnt work either and i didnt expect it to
work. I need the file so i can judge for myself whether my traversal is
correct or not
2) The traversal isnt telling me which node follows the next. I need to
traverse the lattice so that i am able to convert it into a WCN sausage. How
do i do that?

With thanks,
Shamsa.

2)

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

shamsa abid - 2011-09-14

Another thing i need to ask is that i even tried to make the write_lattice
method return me an array of strings containing the lattice file content and
return it as char ** return type.
Now swig eventually makes it a jlong type. But how would that be useful to
access the array in java?

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

Nickolay V. Shmyrev - 2011-09-14

1) I am unable to write the complete lattice file on android, i even tried
giving my hard disk path, that didnt work either and i didnt expect it to
work. I need the file so i can judge for myself whether my traversal is
correct or not

If you open pocketsphinx.log you can easily access error reason. It should
hint you what is wrong and why application can't create a file. The reasons
are always trivial, for example you pointed a wrong directory.

2) The traversal isnt telling me which node follows the next.

You can access this information with ps_latnode_entries/ps_latnode_exits
functions which return iterators over entry links and exit links.

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

shamsa abid - 2011-09-14

First issue is resolved. I actually pushed an empty text file in that
directory and then it worked fine.

For the second case, I was wondering whether ps_latnode_entries and exits
functions are sufficient to traverse a lattice because my ultimate goal is to
create a WCN.
How should i go about pruning the lattice? There dont happen to be any
functions available for that.
Then i also have to merge the repeated nodes, any functions for that?
Or do i have to construct the WCN in memory?

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

I was wondering whether ps_latnode_entries and exits functions are
sufficient to traverse a lattice

They are sufficient

How should i go about pruning the lattice?

/**
 * Prune all links (and associated nodes) below a certain posterior probability.
 *
 * This function assumes that ps_lattice_posterior() has already been called.
 *
 * @param beam Minimum posterior probability for links. This is
 *         expressed in the log-base used in the decoder.  To convert
 *         from linear floating-point, use
 *         logmath_log(ps_lattice_get_logmath(), prob).
 * @return number of arcs removed.
 */
POCKETSPHINX_EXPORT
int32 ps_lattice_posterior_prune(ps_lattice_t *dag, int32 beam);

Then i also have to merge the repeated nodes, any functions for that?

Sorry, I'm not sure what do you mean by that

Or do i have to construct the WCN in memory?

You can use files or work in memory, it doesn't matter.

sphinx lattice to WCN conversion

Speech Recognition Toolkit

Forums

Help

sphinx lattice to WCN conversion document.SUBSCRIPTION_OPTIONS = { "thing": "topic", "subscribed": false, "url": "subscribe", "icon": { "css": "fa fa-envelope-o" } };

sphinx lattice to WCN conversion