Newsgroups: comp.speech
Path: pavo.csi.cam.ac.uk!doc.ic.ac.uk!agate!ames!think.com!spool.mu.edu!umn.edu!myria!dfrankow
From: dfrankow@myria.cs.umn.edu (Dan Frankowski)
Subject: How choppiness affects speech quality
Message-ID: <dfrankow.719564636@myria>
Keywords: speech recognition
Sender: news@news2.cis.umn.edu (Usenet News Administration)
Nntp-Posting-Host: myria.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> <1992Oct10.181427.13448@afterlife.ncsc.mil> <dfrankow.718986383@giga>
Date: Tue, 20 Oct 1992 07:03:56 GMT
Lines: 28

I have checked out several of the references on objective measures of
speech quality that model auditory perception.  They seem to make two
assumptions:

(1) We wish to measure the quality of a coded signal vs. the original
signal, as well as how that is affected by single bit errors.

(2) If there are any time lags, we will properly line up the speech
samples so that they may be properly compared.

Both of these assumptions are the exact opposite of what I want!  In
my case (transmitted digitized speech in real-time over a network), I
don't or care know what the speech quality was before coding.  It is
"coded" when recorded, so on the Sun it is already 8-bit ulaw.  And
my user-level code never sees any bit errors because the transport
layer of the network throws out corrupted packets.  Finally, the
major effect on speech quality in my application seems to be
choppiness and time lags.

Does anyone know of measurements on how choppiness or lags in speech
transmission affect subjectively measured quality?  (Say, the Mean
Opinion Score.)  Any other suggestions would also be welcome.  Saying
how many bytes my application dropped or delayed does not give the
reader a good idea of how good the speech sounded.

Dan
--
Dan Frankowski                dfrankow@cs.umn.edu
