<!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
<meta content="text/html;charset=ISO-8859-1" http-equiv="Content-Type">
<title></title>
</head>
<body bgcolor="#ffffff" text="#000000">
Jean-Marc Valin wrote:
<blockquote cite="mid1130747934.8454.33.camel@localhost.localdomain"
type="cite">
<blockquote type="cite">
<pre wrap="">Add the ability to set adaptation time and min noise duration,
</pre>
</blockquote>
<pre wrap=""><!---->
What do you use that for in practice?
</pre>
</blockquote>
<br>
I use this same ideas:<br>
<br>
(1) As far as "min noise duration", I basically run the preprocessor
VAD inside of some code which continues to call audio "speech" for
about 10 frames past the last time the preprocessor called it speech.
In theory, if the VAD were perfect, I'd be calling 10 extra
(non-speech) frames speech, but in practice, this helps reduce
false-negatives, and only gives my up to 10 extra false-positives at
the end of each spurt. <br>
I think I do this now in app_conference, but not in iaxclient
anymore.<br>
<a class="moz-txt-link-freetext" href="http://cvs.sourceforge.net/viewcvs.py/iaxclient/app_conference/member.c?rev=1.7&view=markup">http://cvs.sourceforge.net/viewcvs.py/iaxclient/app_conference/member.c?rev=1.7&view=markup</a><br>
around line 66.<br>
<br>
<br>
(2) In my distribution of things with the preprocessor, I've slowed the
adaption time (I think this is the same as tom has), because otherwise,
I've gotten both false-negatives on VAD, and the related issue where
the denoiser was removing the speech intonations of particularly
monotonic speakers.<br>
(actually, looking at it now, it seems that the slower adaptation
constant I've used is what's in SVN now).<br>
if (st->nb_preprocess%250==0)<br>
<br>
<br>
-SteveK<br>
<br>
<br>
<br>
</body>
</html>