Newsgroups: comp.speech
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!nntp.sei.cmu.edu!news.psc.edu!hudson.lm.com!news.math.psu.edu!news.cac.psu.edu!newsserver.jvnc.net!newsserver2.jvnc.net!howland.reston.ans.net!nntp.crl.com!pacbell.com!amdahl.com!amd!amd.com!txnews.amd.com!news
From: paul.teich@amd.com (Paul Teich)
Subject: Effects of speech compression on recognition???
Message-ID: <DCnJyE.G7o@txnews.amd.com>
Sender: news@txnews.amd.com
Nntp-Posting-Host: gijoe
Organization: Advanced Micro Devices
X-Newsreader: Forte Free Agent v0.55
Date: Tue, 1 Aug 1995 22:04:16 GMT
Lines: 20

I have found research on WER effects from voiceband compression
codecs, where the recognizer was trained on speech data which had been
compressed then decompressed.

If anyone has WER data where the system was trained on uncompressed
data then run on data which had been compressed then decompressed
(with whatever codec) -- I am interested in understanding any effects.
I know they will be awful; I'd like to quantify how awful the
resulting WER is for this situation.

The application is PCs which recognize speech through a local mic and
over networked links which compress speech to conserve bandwidth.  Is
the problem solely with filters on the front-end, or are two trained
configurations needed for this app?  Anyone tested multiple codecs to
compare WER?

Thanks,
Paul R. Teich
paul.teich@amd.com

