🚀 go-pugleaf

32 bits sounds about right to me for characterizing all the sounds of human
languages.  This is somewhat greater than the number of features proposed in
The Sound Pattern of English, Chomsky and Halle, 1968. The feature system
proposed there is a standard in linguistics.  Since 1968 there have been many
modifications proposed to this system, but none that would change the total of
features greatly.  The features have to do with positions and movements of the
organs of articulation.

Such a system of representation would be appropriate if one wished to
characterize sounds sufficiently well so as to distinguish the pronunciations
of different words for any language, homonyms aside, and also the dialect of
the speaker, and something of the style of speech.  One could not expect to
represent sound well enough to preserve information about sex, age, or, e.g.
presence of sinus infection.

For the most part, the features in this system are binary-valued.  An exception
is the stress feature, but nowadays a commonly held position is that stress is
more appropriately represented by attributing a structure to a string of
sounds.  So perhaps this is not a problem. It was also held by Chomsky and
Halle that other features should have scalar values for detailed
representations, but no grounds were given, and I don't believe it. However
there are some arguments for scalar values in the literature.

It's hard to see any present application outside linguistics for a
text-processing system based on language-universal sound representation, since
one doesn't find texts already represented this way, and there are no devices
available to transcribe speech into such representations.  The work that is
being done on automatic transcription is, so far as I know, parochial and
essentially unprincipled.

*   Greg Lee
*   U.S. mail: 562 Moore Hall, Dept. of Linguistics
*	       University of Hawaii, Honolulu, HI 96822
*     UUCP:	{ihnp4,seismo,ucbvax,dcdwest}!sdcsvax!nosc!uhccux!lee
*     ARPA:	uhccux!lee@bass.nosc.MIL