Newsgroups: comp.speech
Path: pavo.csi.cam.ac.uk!pipex!uunet!kithrup!hoptoad!pacbell.com!iggy.GW.Vitalink.COM!cs.widener.edu!eff!news.byu.edu!gatech!swrinde!sdd.hp.com!zaphod.mps.ohio-state.edu!sol.ctr.columbia.edu!The-Star.honeywell.com!umn.edu!exa!dfrankow
From: dfrankow@exa.cs.umn.edu (Dan Frankowski)
Subject: Re: Objective measure of speech quality
Message-ID: <dfrankow.718581645@exa>
Sender: news@news2.cis.umn.edu (Usenet News Administration)
Nntp-Posting-Host: exa.cs.umn.edu
Organization: University of Minnesota
References: <lels.9.718355427@unpcs1.cs.unp.ac.za> 	<9228109.19223@mulga.cs.mu.OZ.AU> <dfrankow.718561611@myria> <D88-JWA.92Oct8202432@byse.nada.kth.se>
Date: Thu, 8 Oct 1992 22:00:45 GMT
Lines: 31

In <D88-JWA.92Oct8202432@byse.nada.kth.se> d88-jwa@byse.nada.kth.se (Jon Wtte) writes:

>(Dan Frankowski) writes:

>   the person on the other end hears.  What sorts of filters are there
>   that can take x(n) and y(n) and return a quantitative measure of
>   speech quality?

>None, bar a panel of 100 or more test subjects.

>Jon W{tte, h+@nada.kth.se, Sweden, Phone +46-8-107069

Not so.  There has been research in this area.  The method is this:
think up a filter that you think accurately reflects human perception
(first a lowpass filter, then a comb filter to take into account this
and that part of the cochlea, etc.).  Then (here's the good part) take
the measure and test it on samples of speech that have been rated with
the Mean Opinion Score (MOS), which is 100 people listening to it and
rating it excellent (5), good (4), fair(3), poor (2), yuck(1).

If your measure usually gets the same MOS as the 100 people, it's
probably pretty good.  Neat, eh?

I have a few references, if people are interested.  I have also gotten
a few replies to my original post that have been very informative.
Thanks, people.

Dan

--
Dan Frankowski                dfrankow@cs.umn.edu
