Hi Jose,<br><br>This provides a good definition of endpointing as it relates to spoken dialogue systems:<br><br><blockquote style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;" class="gmail_quote">
<table width="617" border="0" cellpadding="0" cellspacing="0" height="173"><tbody><tr></tr><tr>
<td colspan="11" valign="top" width="615" height="215"><font size="2" face="Verdana, Arial, Helvetica, sans-serif">Endpointing
is the process of determining the beginning and the end of speech within an
incoming sample stream. This takes into account whether or not energy is detected
at speech frequencies, the duration of the detected sound and pitch extraction.
The pitch is tracked to help recognise vowels as indications of speech and to
filter out background noise events, such as a door closing. The endpointer can
be tuned to avoid false barge-in, or false-triggers, i.e. cutting off the prompt
when the user did not speak. This can be caused by background noise or prompt
echo. The endpointer can also be tuned to avoid missing leading speech. Syllables,
leading consonants, or in the worst case, everything a caller who speaks quietly
says might be missed. Significant advances have been made in recent years in
echo cancellation and endpointing techniques and these have resulted in better
barge-in performance, and leading speech recognition companies are recommending
the use of barge-in wherever possible. </font></td></tr></tbody></table></blockquote><a href="http://spotlight.ccir.ed.ac.uk/public_documents/Dialogue_design_guide/barge_in_requirements.htm">http://spotlight.ccir.ed.ac.uk/public_documents/Dialogue_design_guide/barge_in_requirements.htm</a><br>
<br>Cheers,<br>Matt<br><br><div class="gmail_quote">2010/6/9 José David Lopes <span dir="ltr"><<a href="mailto:zedavid@l2f.inesc-id.pt">zedavid@l2f.inesc-id.pt</a>></span><br><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;">
Alex,<br>
<br>
Thanks for your quick answer!<br>
<br>
What do you mean by end-pointers? Is it the run mode configuration? I've been looking at the documentation and it is not clear to me what these end-pointers are.<br>
<br>
Jose<br>
<br>
Em 08-06-2010 15:31, Alex Rudnicky escreveu:<div><div></div><div class="h5"><br>
<blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;">
Jose,<br>
<br>
There are actually three different end-pointers in Olympus; these are selectable in (I believe) the configuration for the Audio server. You might try experimenting with alternate ones. Also, it might be worth a try to modify the end-pointing sensitivity parameters.<br>
<br>
Alex<br>
<br>
<br>
-----Original Message-----<br>
From: <a href="mailto:olympus-developers-bounces@mailman.srv.cs.cmu.edu" target="_blank">olympus-developers-bounces@mailman.srv.cs.cmu.edu</a> [mailto:<a href="mailto:olympus-developers-bounces@mailman.srv.cs.cmu.edu" target="_blank">olympus-developers-bounces@mailman.srv.cs.cmu.edu</a>] On Behalf Of José David Lopes<br>
Sent: Tuesday, June 08, 2010 6:56 AM<br>
To: <a href="mailto:olympus-developers@cs.cmu.edu" target="_blank">olympus-developers@cs.cmu.edu</a><br>
Subject: [Olympus developers 233]: Problems with VAD<br>
<br>
I'm working on a dialogue system with a different ASR module. I'm<br>
experiencing some problems with the VAD. This module seems very<br>
sensitive, triggering very easily, and thus forwarding the data coming<br>
from the microphone to the engines even if it is not speech. I was<br>
wondering if I could convert our own GMM models for VAD to the sphinx 3<br>
format. Is there any tool available to convert a standard format (for<br>
instance HTK) to sphinx 3 format? I could also train GMMs using our own<br>
Olympus data. Did anyone did this before?<br>
<br>
Best,<br>
Jose David<br>
<br>
</blockquote>
<br>
</div></div></blockquote></div><br>