<html>
<head>
<meta content="text/html; charset=ISO-8859-1"
http-equiv="Content-Type">
</head>
<body bgcolor="#FFFFFF" text="#000000">
<div class="moz-cite-prefix">You need to run a few sanity checks to
make sure things are working as you expect them to.<br>
<br>
1. Decode 1-best from the HTK lattice WITHOUT rescoring. The
results should be the same as from the HTK decoder. If not there
might be a difference in the LM scaling factor, and you may have
to adjust is via the command line option. There might also be
issues with the CTM output and conversion back to MLF. <br>
<br>
2. Rescore the lattices with the same LM that is used in the HTK
decoder. Again, the results should be essentially identical.<br>
I'm not familiar with the bigram format used by HTK, but you may
have to convert it to ARPA format.<br>
<br>
3. Then try rescoring with a trigram.<br>
<br>
Approaching your goal in steps hopefully will help you pinpoint
the problem(s).<br>
<br>
Andreas<br>
<br>
On 11/22/2012 5:06 AM, Dmytro Prylipko wrote:<br>
</div>
<blockquote cite="mid:50AE235E.7060400@ovgu.de" type="cite">
<meta http-equiv="content-type" content="text/html;
charset=ISO-8859-1">
Hi,<br>
<br>
I found that the accuracy of the recognition results obtained with
HVite is about 5% better with comparison to the hypothesis got
after rescoring the lattices with lattice-tool.<br>
<br>
HVite do not really use an N-gram, it is a word net, but I cannot
really figure out why does it work so much better than SRILM
models.<br>
<br>
I use the following script to generate lattices (60-best):<br>
<br>
<font face="Courier New">HVite -A -T 1 \<br>
-C GENLATTICES.conf \<br>
-n 20 60 \<br>
-l outLatDir \ <br>
-z lat \<br>
-H hmmDefs \<br>
-S test.list \<br>
-i out.bigram.HLStats.mlf \<br>
-w bigram.HLStats.lat \<br>
-p 0.0 \<br>
-s 8.0 \<br>
lexicon \<br>
hmm.mono.list</font><br>
<br>
Which are then rescored with:<br>
<br>
<font face="Courier New">lattice-tool \<br>
-read-htk \<br>
-write-htk \<br>
-htk-lmscale 10.0 \<br>
-htk-words-on-nodes \<br>
-order 3 \<br>
-in-lattice-list srclat.list \<br>
-out-lattice-dir rescoredLatDir \<br>
-lm trigram.SRILM.lm \<br>
-overwrite<br>
<br>
find rescoredLatDir -name "*.lat" > rescoredLat.list<br>
<br>
lattice-tool \<br>
-read-htk \<br>
-write-htk \<br>
-htk-lmscale 10.0 \<br>
-htk-words-on-nodes \<br>
-order 3 \<br>
-in-lattice-list rescoredLat.list\<br>
-viterbi-decode \<br>
-output-ctm | ctm2mlf_r > out.trigram.SRILM.mlf</font><br>
<br>
Decoded with HVite (92.86%):<br>
<br>
<font face="Courier New"> LAB: <A> wie sieht es aus mit
einem weiteren zweitaegigen mit einer weiteren zweitaegigen
arbeitssitzu <br>
REC: <A> wie sieht es aus mit einem weiteren zweitaegigen
in einer weiteren zweitaegigen arbeitssitzu</font><br>
<br>
... and with lattice-tool (64.29%):<br>
<br>
<font face="Courier New"> LAB: <A> wie sieht es aus mit
einem weiteren zweitaegigen mit einer weiteren zweitaegigen
arbeitssitzu<br>
REC: <A> wie sieht es aus mit einen weiteren zweitaegigen
dann bei einem zweitaegigen arbeitssitzung</font><br>
<br>
Corresponding word nets and LMs have been built using the same
vocabulary and training data. I should say that for some sentences
SRILM outperforms HTK, but in general it is roughly 5-7% behind.<br>
Could you please suggest why is it so? Maybe some parameter values
are wrong?<br>
Or should it be like this?<br>
<br>
I would be greatly appreciated for help.<br>
<br>
Yours,<br>
Dmytro Prylipko.<br>
<br>
<fieldset class="mimeAttachmentHeader"></fieldset>
<br>
<pre wrap="">_______________________________________________
SRILM-User site list
<a class="moz-txt-link-abbreviated" href="mailto:SRILM-User@speech.sri.com">SRILM-User@speech.sri.com</a>
<a class="moz-txt-link-freetext" href="http://www.speech.sri.com/mailman/listinfo/srilm-user">http://www.speech.sri.com/mailman/listinfo/srilm-user</a></pre>
</blockquote>
<br>
</body>
</html>