<html>
<head>
<meta content="text/html; charset=ISO-8859-1"
http-equiv="Content-Type">
</head>
<body text="#000000" bgcolor="#FFFFFF">
<div class="moz-cite-prefix">On 4/23/2013 10:22 PM, E wrote:<br>
</div>
<blockquote
cite="mid:8D00EC2D02BB615-71C-6A9E@webmail-d166.sysops.aol.com"
type="cite"><font size="2" color="black" face="arial">Thanks for
the response Andreas.<br>
<br>
I will share my script once its ready. <br>
<br>
<div>This "oracle" WER seems like a very crude way of computing
nbest-error to me. Suppose a reference word is located in [0,
1] seconds, one can look at all the alternatives in the nbest
list (all words that significantly overlap with reference
word) and choose the word that best matches. </div>
<div><br>
</div>
<div>So basically one will extract "most accurate" segments from
each nbest hypothesis in order to get a new "oracle"
hypothesis.</div>
<div><br>
</div>
<div>Do you know if people have done that kind of thing while
computing nbest error?</div>
<div><br>
</div>
<div>Thanks,</div>
<div>Ethan</div>
</font>
</blockquote>
What you suggest is not how nbest WER is commonly defined. However,
taking different pieces from different hypotheses and glueing them
together for an overall better result is the idea behind "confusion
networks" (aka word sausages, or word meshes in SRILM terminology).
<br>
<br>
You can read more about confusion networks at
<a class="moz-txt-link-freetext" href="http://arxiv.org/pdf/cs/0010012">http://arxiv.org/pdf/cs/0010012</a> .<br>
<br>
The nbest-lattice tool in SRILM builds confusion networks from nbest
lists. It also has functionality to compute the lowest WER and
best path through the network.<br>
<br>
Andreas<br>
<br>
</body>
</html>