Newsgroups: sci.lang
Path: cantaloupe.srv.cs.cmu.edu!rochester!cornellcs!travelers.mail.cornell.edu!news.kei.com!news.mathworks.com!newsxfer.itd.umich.edu!chi-news.cic.net!news.midplains.net!gw2.att.com!nntpa!mac-118.lz.att.com!user
From: rte@elmo.lz.att.com (Ralph T. Edwards)
Subject: Re: Power spectrum of phonemes?
Message-ID: <rte-0210951046590001@mac-118.lz.att.com>
Sender: news@nntpa.cb.att.com (Netnews Administration)
Nntp-Posting-Host: mac-118.lz.att.com
Organization: AT&T Bell Labs
References: <AC7F815C966813DB1@yarn.demon.co.uk> <rte-2209951437010001@mac-118.lz.att.com> <DFJtKr.AEI@eskimo.com> <rte-2709951124530001@mac-118.lz.att.com> <DFs1y1.C59@eskimo.com>
Date: Mon, 2 Oct 1995 15:46:59 GMT
Lines: 87

In article <DFs1y1.C59@eskimo.com>, rickw@eskimo.com (Richard Wojcik) wrote:

> In article <rte-2709951124530001@mac-118.lz.att.com>,
> Ralph T. Edwards <rte@elmo.lz.att.com> wrote:
> >In article <DFJtKr.AEI@eskimo.com>, rickw@eskimo.com (Richard Wojcik) wrote:
> >> In article <rte-2209951437010001@mac-118.lz.att.com>,
> >>  [snip]
> >> Psychological phonemes might not have any physical
> >> manifestation in the speech stream.
> >
> >Huh?
> 
> A rather large class of phonological processes have evolved to reduce and
> simplify articulation.  For example, English speakers often articulate a
> 2-syllable word such as "police" /p@lis/ as a single syllable [plis].

Ah, well sure.  You could have been clearer.  I think the problem here is
a typical netnews misunderstanding, where individual words the writer
thinks are clear are mysterious to the reader.

In particular when I say directly comparable I never meant that to mean
completely comparable.  I guess I thought it that was obvious that they
are not completely comparable.  We could make a long list of differences. 
Machine communication ususally uses fixed, uniform symbol times.  Machines
generally have exactly the same number of symbols on both ends.  Correctly
functioning machines do not delete symbols in generation (your point
here).  Usually, the source of error is noise added to the transmission
channel (just one of many sources of error in human communciation).  The
error correction process in machines is one stage of the decoding process,
reducing redundancy placed there by the transmitter according to a precise
algorithm.  Error correction in humans slops over several layers of
decoding, and is imprecise.
And on and on.

> 
> >>  Whether or not you insist on phonemes
> >> having a unique phonetic identity is, again, a matter of theoretical
> >> viewpoint.  
...snip
> 
> All I meant by my statement was that the definition of phonemes has been
> controversial.  That is a matter of historical record.  Not everyone agrees
> (or has agreed) that phonemic distinctions can be neutralized in speech.

Neutralized in speech?  Huh?  (Mysterious)

>  I
> suspect that your difficulty in interpreting my statement may not be that
> we disagree on the essentials, but that you are working off of the term
> "phoneme" as used in the signal processing literature.  If so, then we are
> equivocating on the term and ought to be very careful.  The term "phoneme"
> means something very different in the linguistic literature.
>
 
Well actually it's even worse than that.  Most of my reading has been
in historical or modern languages (amateur's interest, but I've read a lot.)
My knowledge of signal processing is all with machine to machine communication.

>
...
We do not yet know how to use all
> of the different sources of information (pragmatic, semantic,
> morphological, etc.) that humans use to process speech.  Machines just
> don't "hear" phonemes the way humans do.  To give you a better sense of
> what I'm talking about, I'll put up a short paper that Jim Hoard and I
> wrote a few years ago.  It'll be at http://www.eskimo.com/~rickw/bern.html.

Could you mail it?  I can't access it for some reason.  MSWord,
Postscript, whatever.

> 
> BTW, I am not challenging your claim that there is a finite number of
> phonemes for any given person.  I am unhappy with your equating of human
> phonological processing with phonetic segmentation performed by machines.
> I believe that all machine-based algorithms today simply store all of the
> allophonic variation (encoded as "phonemes" [sic]) for words in the
> lexicon.  Humans seem to require considerably less memory to store words,
> because they only need to assign a single phonemic representation to a
> morpheme (ignoring allomorphy) in memory.

I didn't mean to equate them, as I stated above.  There are important
similarities, and important differences.  Certainly in humans the decoding
process is messier, so far somewhat opaque to analysis, and far more error
prone, as we have demonstrated.

-- 
R.T.Edwards rte@elmo.att.com 908 576-3031
