Newsgroups: comp.speech
Path: pavo.csi.cam.ac.uk!doc.ic.ac.uk!agate!apple!mumbo.apple.com!gallant.apple.com!minow.apple.com!user
From: minow@apple.com (Martin Minow)
Subject: Re: information about DECtalk
Sender: news@gallant.apple.com
Message-ID: <minow-171192170923@minow.apple.com>
Date: Wed, 18 Nov 1992 01:26:51 GMT
References: <ulrike.721401334@gmd.de> <minow-101192102931@minow.apple.com> <13566@ecs.soton.ac.uk>
Organization: Macintosh Developer Services
Followup-To: comp.speech
Lines: 49

In article <13566@ecs.soton.ac.uk>, rid@ecs.soton.ac.uk (Bob Damper) wrote:
> 
> In <minow-101192102931@minow.apple.com> minow@apple.com (Martin Minow) writes:
> 
> >-- The letter-to-phoneme rules used in the original DECtalk were based on
> >   Chomsky & Halle, Sound Pattern of English. 
> 
> Not quite.  The *formalism* (i.e.~context-dependent rewrite rules)
> comes from Chomsky and Halle (1968) but not their specific content.
> This formalism was first used for text-to-phoneme conversion by Bill
> Ainsworth back in 1973.  Since then, it has become very popular.

Well, some more history: Sherri Hunnicutt's implementation of the rules
followed Chomsky & Halle fairly closely, although Sherri, Dennis Klatt,
Tony Vitale, and I made some changes here and there based on our testing.
Since I don't have the sources any more, I can only guess that about 75%
of the DECtalk-1 rules closely follow rules in Sound Pattern of English.
Her rules were published in the AJCL microfiche.

Computer-based text-to-phoneme software was also written at the University
of Michigan in 1971 or 1972 (a large Fortran program), but I forgot the
researcher's name; Joyce Freeman, perhaps.

Also, I did my Master's thesis in 1966 on generating surface structure
from deep syntactic structure. That program was written in Snobol-4, ran
rather slowly on an IBM-7090, and was loosly based on some work done
previously at Mitre corporation. It supported both context-sensitive and
transformational re-writing rules. My thesis might be available from the
University of Illinois Department of Computer Science, but I doubt that
it's still interesting.

Regarding text-to-speech testing: we did preliminary testing of DECtalk
using a large (over 100,000 words) dictionary that had been
hand-phonemicized.
A program ran the dictionary words through the DECtalk rule-set then
compared them against the dictionary definition (ignoring uninteresting
differences such as unstressed-i versus shwa). Adjusting the rules changed
the set of erroneous words and, eventually, the remaining words were put
into the dictionary.

After DECtalk was reasonably complete, and during the development of
DECtalk-III (the small-circuit board version), we made extensive use
of a user-based testing lab at the University of Indiana under Prof.
Pisoni. Testing using "live" subjects proved extremely useful to
guage understandability -- something that is not identical to phonemic
correctness, of course.

Martin Minow
minow@apple.com
