minimal pairs are not always there

[Robert Whiting]

On Wed, 26 Apr 2000 Gábor Sándi <g_sandi at hotmail.com> wrote:

>[ Moderator's note:
>  Robert Whiting's posting quoted below was in response to, and
>  quoted from, a previous message from John McLaughlin dated 13
>  Apr 2000.  I have added the proper attribution where needed.
>  --rma ]

>On Saturday, 22 April, 2000 Robert Whiting
><whiting at cc.helsinki.fi> wrote:

>Subject: RE: minimal pairs are not always there

>[ JMcL wrote:]

>>> [Robert Whiting]
>>> I would say that even a minimal pair is not a sufficient
>>> condition to establish two sounds as separate phonemes.

>I am entering this debate somewhat late, but I do have strong
>opinions on it.

>In my view, the main purpose of phonemic analysis is to provide
>for an unambiguous way to describe the pronunciation of every
>utterance in a language.

Yes, well, this is what we have the IPA for.  But you may have
noticed that IPA is the International PHONETIC Alphabet, not the
*International PHONEMIC Alphabet.  You are simply confusing
phonemics with phonetics.  Phonemic transcription (/.../) is
neither phonetically unambiguous nor necessarily phonetically
accurate.  Phonetic transcription ([...]) is necessarily (or at
least to the best of the transcriber's ability) an accurate
mapping of the pronunciation of every utterance in a language.
But phonetic transcription doesn't care what the phonemes are; it
records the actual sounds (phones) of the utterance.  As a simple
example, English 'pin' is transcribed phonemically as /pin/ but
phonetically as [p'in] or [p^hin] because the initial /p/ is
aspirated (followed by a puff of air).  The consensus is that
aspirated and unaspirated /p/ in English are simply allophones of
a single phoneme (although a case could be made that they
aren't).  Now consider 'spin', which has the unaspirated
allophone of /p/ (or /P/ if you are into archiphonemes).
Phonemically it is /spin/ but phonetically you could most
accurately represent it as [sbin] (the puff of air of the
aspirated consonant delays the onset of voicing; without the puff
of air, voicing of the vowel sound begins before the consonant is
finished, making the consonant seem voiced).  So while the
phonetic transcription may be accurate, it is at odds with the
phonemic transcription.

Whenever you have allophones of a single phoneme, phonemic
transcription is not necessarily an accurate phonetic mapping of
an utterance since each phoneme can have several members.
Phonemic and phonetic transcriptions do not have to look very
much alike.  This is why we have a provision for both phonetic
and phonemic transcriptions.  A phonetic transcription is an
accurate mapping of the pronunciation of an utterance.  A
phonemic transcription is an abstraction.

This is because phonemes themselves are abstractions.  Phones are
realia.  They exist in the real world.  They can be measured and
identified with a voice spectograph.  Phonemes exist in the minds
of the speakers of a language (and for the most part, they don't
even exist there consciously).  Two (or more) phones may be
allophones of the same phoneme in one language but separate
phonemes in another.  You can't tell from looking at a voice
spectograph whether a phone is a phoneme or not.  You can't even
tell what the phonemes of a language are by asking its speakers,
because they won't be able to give you a list of the phonemes of
their language.  Native speakers learn their phonemes through
contrasts.  The never sit down with a list of the phonemes and
memorize them the way the memorize multiplication tables.  You
can only tell what the phonemes of a language are by analyzing
the speech of its speakers and seeing which sounds contrast and
under what circumstances.

>Therefore if there is a single pair of words distinguished by
>the presence of sound A in one and sound B in the other (this is
>the definition of "minimal pairs"), this should be sufficient to
>establish a phonemic difference. In any dialect of English where
>"either" may be pronounced /i:dh at r/ (@ stands for the schwa), the
>existence of the minimal pair either/ether is then sufficient to
>establish the existence of separate phonemes /dh/ and /th/.

This is like saying that [x] and [ç] (the ich-laut) must be
separate phonemes in German because you can find "minimal pairs"
with these two sounds ('Kuhchen' [ku:çen] "little cow" and
'Kuchen' [ku:xen] "cake" or 'Tauchen' [tauçen] "little rope" and
'tauchen' [tauxen] "to dive, submerge").  This is just not so.
The presence of these sounds can be explained by rule (even if it
isn't a phonological rule).  I expect that most native speakers
of German would not accept these as "minimal pairs" even if they
are by your definition.

And a single "minimal pair" doesn't establish separate phonemes;
it establishes a contrast between two segments.  If one accepts
'ether' and 'either' as a minimal pair, this proves only that
[th] and [dh] contrast.  It doesn't prove that [th] and/or [dh]
are not allophones of /t/ and/or /d/.  See below.

Perhaps we can eliminate some of the discord about whether
minimal pairs define phonemes by refining the definition of
minimal pair, rather than refining the definition of phoneme.
The problem, in my view, stems from defining phoneme in terms of
meaning.  Unfortunately, because of the way language works (as a
system that expresses meaning through sound), it is very
difficult to avoid tying meaning to sound definitionally at some
point.  But phonemes are not about differences in meaning;
phonemes are about being able to make arbitrary contrasts that
result in differences in meaning.  Differences in meaning may be
a result of phonemes.  Phonemes are not a result of differences
in meaning.  The key word here is "arbitrary".  The relationship
between sound and meaning is arbitrary.  If two words differ by
only one segment, the differing segments are different phonemes
only if their presence in the words is arbitrary.  If the sounds
are required to be in those words by some rule, the distinction
can't be phonemic.  To say that words with different meanings
must have different phonemes is to require that homonyms have
invisible (inaudible?) phonemes that provide the difference in
meaning.

So let us keep the meaning of phoneme as "the smallest unit of
sound that can make a difference in meaning" (although I still
don't like this, because in practice any difference in sound can
make a difference in meaning, but any difference in sound is not
necessarily a different phoneme) with the proviso that phonemes
do not make any *specific* difference in meaning.  A phoneme
should be semantically empty.  If it is not, then it is no longer
(just) a phoneme.  Phonemic contrasts must be arbitrary.  If we
have the words 'big' and 'pig' and 'fig', these are arbitrary
contrasts.  Nothing requires the phonemes /b/ or /p/ or /f/ to be
in any of these words. We can check this by comparing other words
like 'bad', 'pad', and 'fad' or 'ban', 'pan', and 'fan' and so
on, and seeing that there is nothing about any of the phonemes in
these words that tells you anything about their meanings or
anything about their meanings that tells you what particular
phonemes have to be present.  To put this another way, ask
yourself a question:  If the phoneme /b/ occurs in an English
word, what does that tell you about the meaning of that word?
The answer should be:  "absolutely nothing."

In the pair ether/either, 'ether' belongs to a class of words
where the pronunciation of <th> as [th] is required (by rule
based on a clearly established pattern [Greek loans originally
containing theta]); 'either' belongs to a class of words where
the pronunciation of <th> as [dh] is required (by rule based on a
clearly established pattern [native words with an intervocalic
spirant where the original voiceless spirant was voiced by
rule]).  Since the respective pronunciations of these words are
required by rule, this cannot be a phonemic distinction because
the choice of these sounds is not arbitrary.  If 'ether' had [dh]
and 'either' had [th], this would be a phonemic distinction
because the opposition would not be predictable.  Saying that
native speakers don't know that 'ether' is originally from a
Greek word with theta doesn't make any difference.  There must be
some synchronic rule that forces this [th] pronunciation or else
the pattern wouldn't be there in the language synchronically.

Now it might be argued that [th] and [dh] don't make any
*specific* difference in meaning in these two words and therefore
the contrast should be considered phonemic.  And it is true that
'ether' could mean anything and 'either' could mean anything
else.  But so long as 'ether' comes from a Greek word that had
theta in it and 'either' is a native word that had an
intervocalic /th/ when the voicing rule operated, then 'ether'
would *have* to be pronounced with [th] and 'either' would *have*
to be pronounced with [dh].  There is nothing arbitrary about the
choice of [th] and [dh]; these pronunciations are required by
rule, therefore the contrast is not phonemic.

>Of course, other pairs like thy/thigh, this'll/thistle etc. will
>reinforce this analysis.

Indeed they will.  Whenever the distribution of the sounds can be
explained by rule, the distinction is not phonemic.  The point
about phonemes is that a speaker can use them to make a
distinction in meaning.  When a speaker is required to use a
certain sound in a certain place (for reasons other than simply
to effect the desired meaning) then that usage is not phonemic
(for all that it may create a different meaning).

>You may of course not accept my approach, but I do not see the
>utility of a phonemic system that cannot uniquely map the
>pronunciation of every word in a language.

But this is not the purpose of a phonemic system.  Phonemic
systems do not uniquely map the pronunciation of every word in a
language.  Again, you are confusing phonemics with phonetics.
Phonemic analysis accounts for the different phones used in a
language by assigning them to phonemes.  And this is done by
determining which phones contrast and which do not.  But a
phonemic inventory is coextensive with a phonetic inventory of a
language only if there are no allophones.

>Imagine writing a description of the English sound system, say
>of the "General American" dialect. If you don't accept the
>phonemicity of /dh/, you will presumably provide a list of eight
>fricatives:

>/f/ /v/ /th/ /s/ /z/ /sh/ /zh/ /h/

>Next you will say that some of these come in voiceless/voiced
>pairs: /f/ - /v/, /s/ - /z/ and /sh/ - /zh/. One (/h/) is always
>voiceless. Finally, there is a curious phoneme /th/, which is
>voiceless initially (except in function words) and in word-final
>position in nouns (bath) and adjectives (uncouth), but voiced
>between vowels (except when not), after consonants (further), and
>in word-final position in verbs (bathe). Words of foreign origin
>(as if native speakers cared) would have their own rules (Athens,
>anthem). No doubt other sub-rules could be added, ad infinitum.

>Now why would this analysis be superior to one that said that
>English had nine fricatives (/f/  /v/  /th/  /dh/  /s/  /z/  /sh/
>/zh/  /h/ )?

Because it is more accurate and accounts for all of the evidence?

Besides, I don't deny that English has nine fricatives.  You can
prove it with a spectograph.  What I deny (or at least can't find
evidence to prove) is that all these nine fricatives (or spirants
as I prefer to call them) are phonemically distinct.  The fact
that certain sounds occur in a language doesn't ipso facto prove
that they are all phonemes in that language.  English (and most
other languages as well) has lots of sounds that aren't phonemes
(or, rather, that are different from, but not distinct from,
other members of the same phoneme).

I'm surprised that you haven't included /x/ as an English
fricative.  Although once present and then lost, is has surely
been borrowed back in words like 'ach', 'loch' and names like
'Bach'.

>Here you need no distributional rules, just a specification of the
>articulation of each phoneme: voiceless labiodental fricative etc.
>[snip]

I fail to see the point of this.  You can describe the
articulation of each sound regardless of whether it is a phoneme
in any particular language or not.  [th] is still a voiceless
alveolar (dental) fricative and [dh] is still a voiced alveolar
(dental) fricative regardless of whether they are separate
phonemes or not.  They will be pronounced the same way whether
they are separate phonemes or allophones of the same phoneme.
Phonemicity doesn't have anything to do with how a sound is
pronounced.  Two sounds can be pronounced differently (have
different features) and still not be phonemically distinct.  Two
sounds can be phonemically distinct in one language and not
phonemically distinct in another, but they will still be two
sounds.  Consider bright /l/ and dark /l/:  allophones of /l/ in
English, but separate phonemes in Russian.

Now a distinctive feature analysis may make it easier to check
sounds cross-linguistically, but for a single language, it makes
little difference whether you identify its sounds through
features or as phonemes and their allophones.  Many sounds have
both distinctive features and redundant features.  Redundant
features belong to allophones of sounds whose distinctive
features contrast (in at least one feature) with all the other
phonemes in the language.  For example English /k/ has at least 6
allophones:  a palatal, a labial, a neutral, with aspirated and
unaspirated variants of each.  But for purposes of phonemic
analysis, English /k/ is simply a voiceless velar stop.  The
redundant features that distinguish its allophones are just not
considered in phonemic analysis.

>> No problem.  I have no vested interest in any theory that
>> either requires or doesn't require [th] and [dh] to be separate
>> phonemes. I'm just looking to find out what the evidence is and
>> how the evidence proves it one way or the other.

>[ JMcL wrote:]

>>> The evidence for establishing /th/ and /dh/ as separate
>>> phonemes is no worse than that for establishing /zh/, /ng/, and
>>> /oj/ as phonemes (depending on whether or not one considers
>>> diphthongs to be on the same footing as other phonemes in the
>>> language).

>> I think it is quite a bit worse.  How many of /zh/, /ng/, or
>> /oj/ occur only in certain classes of words or only as
>> morphophonemic alternants?  Show me that in all the words where
>> /zh/ occurs that /zh/ limits or restricts the meanings that it
>> can have and I will grant you the point.  It is said that /ng/
>> only occurs in word final position, but even this is not true
>> (compare 'finger' - 'singer').  These may be difficult to
>> establish as phonemes, but there is solid evidence: places where
>> these sounds provide the only contrast and cannot be predicted by
>> rule.  Show me the same for [dh] and I will grant you the point.

>[GS]

>Once you accept nonphonetic conditioning factors, there is no
>end to the elimination of phonemes from a system.

Of course there is an end.  This is just hyperbole.  There is no
way to eliminate /p/, /b/, /t/, /d/, /k/, /g/, /f/, /v/, /s/,
/z/, /th/, /sh/, /m/, /n/, and /h/ as phonemes in English.  It is
rather the other way around:  if you eliminate all nonphonetic
conditioning factors and say that all sounds that are potentially
distinctive are phonemes, there is no end to the number of phonemes
in a language.

Phonological purists always insist that morphology can't affect
phonology (i.e., no nonphonetic conditioning factors) and the
next thing you know they are talking about the phonological
effects of morpheme boundaries and word boundaries.  In fact, it
is almost impossible to discuss phonology without these concepts.
But morpheme boundary and word boundary are not phonological
conditions, but morphological ones.  Somehow they seem to get
smuggled into phonology.  So it is not a matter of "no
nonphonetic conditioning factors," it is a matter of how much
morphological conditioning you are willing to allow into
phonology.

My own answer to this tends to be "as much as you can."  Language
is a fully integrated system.  It expresses meaning (morphemes)
through phonological form (phonemes).  So in my view it is very
difficult to completely separate morphology from phonology.
Systems engineers know that when you start mucking around with
one part of a system, you are likely to cause unexpected havoc in
some other (apparently unrelated) part of the system.  Software
developers are intensely aware of this; they just can't explain
it to judges, apparently. :)

>Take two of your examples above:

>/ng/ - not a phoneme. We can analyze it as the allophone of /n/
>before /k/ and /g/. /g/ is then dropped ("zero allophone" - why
>not?) in final position after /n/, and before the derivative
>suffix -er (as in "singer"), although not before the comparative
>-er (as in "longer"). There remain some exceptions, like the name
>of my favourite Canadian city (Vancouver is pronounced
>/vaenku:v at r/ by locals), but this is clearly a loanword, or
>(maybe) two morphemes: Van Couver.

Well, yes, we could analyze it this way (with some
modifications), but then we would have to have a silent phoneme
in the language, because you can't just elide final /g/.
Phonemically, its place has to be held by a marker of the zero
allophone of /g/.  Otherwise, in pairs like 'run', 'rung'; 'sin',
'sing'; and especially 'sinner', 'singer' you would have [n] and
[ng] contrasting (and we all know that allophones can't contrast
in different words).  Therefore 'sinner' would have to be
analyzed as /sin at r/ and 'singer' would have to be analyzed as
/sing0(g)@r/ (where 0(g) represents the zero allophone of /g/) or
else you would just have /sin at r/ contrasting with /sing at r/
suggesting that [n] and [ng] are not allophones.  Of course, you
will also have a different syllable structure (si'-n at r vs.
sing'- at r) because [ng] can't occur in syllable initial position.
This doesn't affect [n] / [ng] contrasts in final position,
however, so we have /r at n/ and /r at ng0(g)/.

One of the modifications that it seems to me would have to be
made is that g --> 0 / ng ___# has to be the next rule applied
after n --> ng / ___ [velar].  This gets rid of the complication
of saying that /g/ goes to zero after [ng] in final position and
before the agentive suffix '-er'.  Stem-final /g/ goes to zero
immediately and (nearly) all derivative suffixes (agentive '-er',
'-ing', '-ish', '-ly', '-y', etc.) are added to this base.  All
that is needed then is a rule that triggers the restoration of
/g/ before the comparative '-er' (and for completeness, the
superlative '-est'):  0(g) --> g /ng ___ [COMPARATIVE,
SUPERLATIVE].

So it is not too difficult to account for the distribution of [n]
and [ng] in English with synchronic (generative) rules and
treating [ng] as an allophone of [n].  The only real problem is
that it requires a silent phoneme (a stem-final zero allophone of
/g/).  There is nothing particularly wrong with this (as someone
pointed out, French does a lot of this, and John McLaughlin has
pointed out that Numicists do it for Comanche).  Furthermore, it
is not a particularly difficult phoneme for speakers to keep
track of:  wherever you have a bare [ng] (i.e., without a
following velar), it is followed by a zero allophone of /g/.

There is another problem for phonological purists, because the
initial assimilation rule obviously doesn't operate across
morpheme boundaries (compare 'unclear', 'unclean', but 'uncle')
and sometimes even across perceived morpheme boundaries (e.g.,
'mangrove', 'mongoose', your 'Vancouver', 'vanguard', and
sometimes even 'vanquish' [analogy?]).  This simply means that
the original assimilation rule has to be marked [+root] as part
of the environment.  And the zero allophone of /g/ rule probably
doesn't work in New York City because too many people live on
"Long Gisland." :)

All things considered then, it is probably simpler to operate
with /n/ and /ng/ as different phonemes because this does not
require the zero allophone of /g/ to maintain their allophonic
status *and* there is no indication of a morphophonemic
alternation between [n] and [ng] (there is no connection between
pairs like 'bun' / 'bung', 'pin' / 'ping', 'thin' / 'thing',
'clan' / 'clang', etc., etc.).  /ng/ is frequently used in
expressive and imitative words (e.g., 'ding', 'ping', 'zing',
'ring' 'clang', 'twang', 'bong', etc. -- even
cross-linguistically:  'gong' is a loan from Malay, and the
Sumerian word for "harp" was 'balang'), but not exclusively so
(e.g., 'thing', 'lung' 'rung' [noun], 'strong', 'among', 'gang',
and, most commonly, the derivative suffix '-ing').

But the analysis does show that it is possible to analyze two
sounds both as separate phonemes and as allophones of a single
phoneme, accounting for all the observed data in each case.  In
such instances we usually resort to Occam's Razor for a
resolution.  In this case we accept the separate phoneme theory
because the allophone theory requires an additional entity (a
zero allophone of /g/).  But this does not necessarily mean that
this is the correct solution.  Occam's Razor is just a heuristic.
It tells us that the simpler solution (that accounts for all the
evidence) is to be preferred, but it doesn't tell us that it has
to be correct.  The simple solution is more likely to be correct,
but unlikely things do happen.

>/zh/ - who needs it? Using the logic of reductionism, we shall
>elaborate a set of rules to account for this sound:

This is an oxymoron.  Reductionism doesn't elaborate rules; it
simplifies them.

>(1) It is an allophone of the phoneme /j/, occurring in
>loanwords from French (genre, garage, mirage).

First, let's get our terminology straight.  What we will call the
phoneme /j/ is j with hachek [dzh] and not to be confused with
IPA j which is English y.  As such, /j/ is the voiced counterpart
of /ch/ (c with hachek [tsh]).  Now to facts. The sound [zh]
occurs predominantly in French loans borrowed after about
15-1600, but it also appears in some loans from Slavic; I
remember at least 'muzhik' "peasant" from Russian, but there are
probably others.  But the thrust of the statement is correct:
[zh] in this environment is found only as a received
pronunciation in words of foreign origin.

>The phoneme /j/ in native words like 'edge' /ej/ remains [j],
>whereas this pronunciation does not occur in loanwords.

Sure it does.  It is found in words from Japanese ('jinricksha',
'judo', 'jujitsu' [not to mention 'Japan']); Arabic ('jihad',
'jinni' ['genie']); and many others.  But most of the loanwords
with this pronunciation come from French.  In speaking of loans
from French, the pronunciation of this segment depends on when
the word was borrowed.  Words borrowed before about 15-1600 will
have [j (dzh)]; words borrowed later will have [zh].  This
parallels exactly the treatment of the voiceless counterparts of
these sounds.  Words borrowed before the sound change in French
will have [ch (tsh)] ('chief', 'chandler').  Words borrowed later
will have [sh] ('chef', 'chandelier').  This last change does not
cause any controversy in English because /s/ and /sh/ had already
been established as separate phonemes.  However, the
pronunciation of <ch> as [sh] makes it easy to pick out French
loans of the last 4-500 years.  But 'chief' never had a
pronunciation with [sh] in English.

>Where it does, one may always set up a new word category:
>loanwords that have been fully assimilated to the sound structure
>of English. This will take care of troublesome words like jet,
>gene and magenta.

No, the vast majority of loans from French with [j] never had a
pronunciation of [zh] in English.  This is a matter of a sound
change in French, not of nativization of the sound in English.
So your statement is a distortion of the evidence, especially
of the evidence you have offered.

Of the three examples given, 'jet' is in a different category.
I'm not entirely sure about this, but I seem to remember that _j_
from Latin _j_ went to [zh] later than _ge_ from Latin _ge_.
Perhaps someone can help out on this.  I would need a philologist
of French to verify this, but I expect that 'jet' never had a
pronunciation with [zh] in English despite the fact that it
wasn't borrowed until the 17th century.  Contrast with 'jeté'
[zh at te:] (a technical term in ballet, where unnativized French
terms abound), a 19th century borrowing.  Words like 'joy'
'joust' 'judge', 'jury', and so on, never had a pronunciation
with [zh] in English.  Again, contrast 'juggler' (14th cent.;
with [j]) with 'jongleur' (18th cent.; with [zh]).

The second example is a modern coining based on the suffix
'-gen', as in 'pathogen', 'halogen', 'oxygen', 'hydrogen'.
'Gene' never had a pronunciation with [zh] in English.

The final example is Italian, not French.  The <ge> in 'magenta'
is exactly parallel to its voiceless counterpart <ce> in Italian
loans such as 'cello' or 'vermicelli', regularly pronounced with
[ch].  'Magenta' never had a pronunciation with [zh] in English
(except for those who give it a French pronunciation as an
affectation).

So there are two strata of French loans in English:  the earlier
will have <j> or <g> and [j] ('jet', 'jelly', 'jaunty',
'juggler'; 'gender', 'gelatin', 'genteel'); the later will have
<j> or <g> and [zh] ('genre', 'gelée', 'gendarme'; 'jeté',
'jongleur', 'jeunesse').  The earlier borrowings never had a
pronunciation with [zh] in English.

Now despite your poor choice of examples, there is a process of
nativization going on with recent French loans.  To my knowledge,
'ménage' has only [zh] but 'menagerie' can have either [j] or
[zh].  Similarly 'mélange', 'garage', 'genre', 'gendarme', etc.,
can have either.  The pronunciation of 'garbage' with [zh] is an
affectation; apparently French garbage is considered to have more
class than English garbage. :)  But it does illustrate the point
that the pronunciation with [zh] is considered by speakers to be
connected with French and that a native (or nativized) word can
be "Frenchified" by using this pronunciation..

>(2) In words like azure, seisure and invasion, /zh/ is derived
>from the sequence of phonemes z + y. (/azyu:r/, /si:zy at r/ and
>/inveyzy at n/). Come to think of it, didn't Chomsky and Halle
>analyze English along these lines, to get back at those dreadful
>structuralists of the 50's?

I think that Chomsky and Halle were trying to show that English
pronunciation and English orthography were not so disparate as
generally assumed (as an antidote to Shaw, etc.) and that English
spelling provides a great deal of information to the reader that
the pronunciation has lost.

But the analysis of this instance of [zh] is quite correct.  It
is a palatalized allophone of [z] before a palatal ('confuse'
[z], 'confusion' [zh]) parallel with the palatalized allophone
[sh] of [s] in the same environment ('confess' [s], 'confession'
[sh]).

This is quite regular and very common cross-linguistically:
consonant + palatal --> palatalized consonant.  In short, the
palatal, originally associated with a (usually front) vowel, is
transferred to become associated with the preceding consonant.
Again, [sh] in this environment is a palatalized allophone of
[s], but since /s/ and /sh/ are already considered phonemes
('sore' / 'shore', 'seat' / 'sheet'), the rule (concept) "once a
phoneme, always a phoneme" causes it to be considered as a
phoneme even in this environment where it is clearly an
allophone.

Many people then consider that since /sh/ is a phoneme in this
environment then [zh] must also be a phoneme.  They then collect
numerous examples of minimal pairs and near minimal pairs to show
that /sh/ contrasts with [zh] in the same environment
('Confucian' / 'confusion', 'kosher' / 'closure', 'vicious' /
'vision', 'pollution' / 'collusion') and then claim that this
proves that [zh] is a phoneme because it contrasts with /sh/.

But as Ante Aikio has quite correctly reminded us on Wed, 26 Apr
2000:

   1. One minimal pair, such as above, does not establish
   phonemes - it establishes an opposition in a non-contrastive
   environment. A phoneme is the sum of its oppositions: what
   makes something a phoneme is that it is in opposition to -all-
   the other phonemes. You can establish phonemes only in the
   context of the whole phonological system.

To which I would add that a hundred minimal pairs, or a thousand,
that all express the same contrast (e.g., [sh] and [zh]) does not
establish phonemes, only a contrast between the two particular
segments involved.

This is where the definition of "phoneme" that I use comes into
play.  The definition that I use is that a phoneme is the
smallest unit of sound recognized as distinct from all other such
units by the speakers of a language.  This definition has a
number of disadvantages.  Namely, it does not account for
suprasegmental phonemes and it also defines the concept in terms
of itself.  Its main advantages are that it keeps "meaning" out
of the definition and it stresses the psychological nature of the
phoneme.  And it tells you how you go about identifying phonemes.
To be a phoneme, a sound must contrast with *all* other phonemes
in the language.  If it contrasts with all other phonemes except
one, it is not a phoneme, but an allophone of that one (unless
the two sounds are so distant phonetically that they simply
cannot be considered allophones [like English /h/ and /ng/]).

I consider this a conceptual definition; it tells you what
phonemes are, as opposed to the functional definition ("a phoneme
is the smallest unit of sound that can make a difference in
meaning") which tells you what phonemes do.  Phonemes can make
differences in meaning.  But you can have differences in meaning
without having different phonemes (homonyms).  Therefore, a
difference in meaning is not a sufficient condition for insisting
on the presence of different phonemes.  And a minimal pair is
neither a necessary nor a sufficient condition for establishing
different phonemes.

What is necessary is to show that a sound is a phoneme is to show
that it contrasts with every other phoneme in the language.  And
this is where Ante Aikio seems to have forgotten what he made so
clear in his first point when he wrote in his second point:

   2. It must be once again stressed that for many languages it
   is impossible to establish all the contrasts using minimal
   pairs - this is precisely the case with English /sh/ vs. /zh/,
   which was discussed earlier. The phonemes contrast with each
   other even though there are no minimal pairs, because their
   distribution cannot be accounted for with a rule. I.e., they
   show contrast in non-contrastive environments (vicious vs.
   vision etc.).

The contrast of /sh/ and [zh] does not prove that [zh] is a
phoneme.  As Ante Aikio made clear in his first point, it only
proves a contrast with /sh/.  It cannot be proved that [zh] is
a phoneme unless it can be shown that [zh] contrasts with *all*
other phonemes (or is so phonetically distant from some phoneme
for which a contrast cannot be shown that it cannot be considered
an allophone of that phoneme).  Particularly, it is crucial to
show that [zh] contrasts with /z/.  Because if [zh] is an
allophone of /z/ then [zh] will contrast with all other phonemes
except /z/ because /z/, as a phoneme, will contrast with all
other phonemes and so, therefore, will all of its allophones.
And /z/ and [zh] are so close phonetically that the second
criterion does not apply.

Now it has been correctly pointed out that [zh] in English has
two origins:

   1) It is a received pronunciation in loanwords.  Such
      loanwords are predominantly from French where the
      pronunciation is primarily a result of first palatalization
      of /g/ to /j/ ([dzh]) before [e] and then unpacking of the
      complex sound ([dzh] > [d]+[zh]) and elision of the initial
      [d].  This development took place in French, not in
      English where the phoneme /j/ is still found in both native
      words ('hedge') and borrowed words ('judge').

   2) It is a palatalization of /z/ before palatal vowels (z -->
      zh / ___ [palatal]).  As such, it is clearly an allophone
      of /z/.

Since zh(2) is an allophone of /z/ it cannot contrast with /z/.
It can, however, contrast with /j/ (and, presumably therefore,
with its allophones)  Since we have 'legion' ([j]) and 'lesion'
([zh(2)]) this contrast is validated.  There is also one
pronunciation of 'azure' (I don't know the pronunciation that you
give:  /azyu:r/; I know [æzh at r] and [@zhUr], the first with accent
on the first syllable, the second with accent on the second
syllable) that shows this contrast:  'adjure' [@jUr] (this also
contrasts with [sh]:  'assure' [@shUr]).

On the other hand, zh(1) can contrast with /z/.  Thus we have
'ruse' [ruz] and 'rouge' [ruzh] (although 'rouge' is also
realized as [ruj] in some dialects or by some speakers) and
'lose' [luz] and 'luge' [luzh] (also 'loser' [luz at r] and 'luger'
[luzh at r]).

So we are faced with the fact that zh(1) (as an allophone of /j/)
can contrast with /z/ and zh(2) (as an allophone of /z/) can
contrast with /j/.  But to my knowledge, zh(1) does not contrast
with /j/ and zh(2) does not contrast with /z/.  Thus the question
becomes:  are we justified in considering identical allophones of
different phonemes as phonemes in their own right?  If the answer
is yes, then [zh] is a phoneme in English; if not, not.

In Akkadian, there are three phonemic short vowels (a, i, u) and
three corresponding long vowels.  There is also a short and long
e, but this is not considered phonemic because it is always an
allophone of either a or i.  Now it is possible to find "minimal
pairs" in which e (as an allophone of a) contrasts with i and in
which e (as an allophone of i) contrasts with a.  Thus it is
possible to show four different contrasts, but not that there are
four different vowel phonemes.  I would say that the case of
English [zh] is entirely similar.

On the other hand, Akkadian being a dead language, the grammar
was explicated by philologists rather than by linguists.  A
linguistic approach might reach different conclusions.

>With sufficient ingenuity, I am sure we can come up with rules
>to eliminate other phonemes from English: /v/ and /@r/ (bird,
>fern, word) come immediately to mind.

With sufficient ingenuity, we can come up with anything we want.
The question is, will anyone believe it.

>My main point, I hope readers realize, is that this kind of
>analysis is not very helpful. A simple description of the
>phonemic structure of a language should account for all
>differences that are potentially distinctive, irrespective of the
>morphology, semantics or etymological provenance of the words in
>question.

And it is not very helpful to say that every different sound in a
language is a phoneme.  For any different sounds in a language
are potentially distinctive.  Whether they are actually
distinctive depends on how the speakers of the language use them.
If they aren't used to make arbitrary contrasts, then they aren't
phonemes (in that particular language).

>It is a noteworthy fact that there tend to be correlations
>between certain phonemes and certain grammatical and semantic
>criteria. In English, /dh/ starts many function words and /th/
>never does - so what? The phoneme /h/ (or the cluster /hw/,
>depending on the dialect) starts many interrogatives (where,
>which, what, when) while /k/ never does - is this a reason to
>bunch them together under one phoneme?

This is mere sophistry.  The one situation has no relevance to
the other.  To say that it does may make it look like a valid
argument to anyone who doesn't know anything about historical
linguistics, but it is easy to refute by anyone who does.

The split between initial [th] and [dh] is quite recent in
English.  There can be no question that [th] and [dh] were
originally the same phoneme.  Initial /th/ is the normal Germanic
reflex of PIE *t as captured in Gothic.  English has preserved
/th/ while German has generally shifted to /d/.  Thus:

     PIE       Gothic            English           German

    *ter-      thaurnus          thorn             Dorn
    *trei-     thrija (neut.)    three             drei
    *tr.s-     thaurstei         thirst            Durst
    *tar-      thar              there             dar, da
    *tu:-      thu               thou              du
    etc        etc               etc               etc

By comparison, /hw/ is the normal Germanic reflex of PIE *k^w
(just as /h/ is the normal Germanic reflex of PIE *k' [palatal
k]).  All the interrogatives in /hw/ (English <wh> or <h>) are
derived from the Teutonic pronoun base *hwo- / *hwe- (presumably
phonetically [x^w]) which is derived from PIE *k^wo- / *k^we-.

So the reason that there are no interrogatives beginning with /k/
in English is that there is no initial PIE *k^w left in English.
Any English initial /k/ has come from some other source (normally
shifted from PIE *g or *g^w; otherwise borrowed from a language
that has preserved PIE *k^w like 'quart(er)', or from an
unrelated language ['Quetzalcoatl', 'Kwakiutl').  That is why no
one claims that modern English /hw/ and modern English /k/ are
the same phoneme.  They aren't and they never have been.  English
/hw/ and PIE *k^w are the same (i.e., one has developed into the
other), but there is no initial /k/ in English that has been
inherited from PIE *k^w or *k'.  Therefore saying that [th] and
[dh] can't belong to the same phoneme in English because /hw/ and
/k/ aren't the same phoneme in English is a completely spurious
argument and I can't imagine why anyone who has ever heard of
Grimm's Law would put it forward.

<snip>

>[ JMcL wrote:]

>>> [Me]
>>> The distribution of /th/ and /dh/ cannot be determined by the
>>> assignment of a PHONOLOGICAL rule.

>> Let me see if I have your take on this straight.  Are you
>> saying that allophones automatically become separate phonemes
>> when the phonological conditioning environment that maintains
>> their allophonic identity is lost?  And that they are phonemes
>> even if they never contrast in an environment that can't be
>> predicted, so long as the basis for predicting the environment is
>> not phonological?

>[GS]

>Yes, that's how it should be. Once the conditioning factor is
>lost, we have separate phonemes. German (+Swedish, Danish and
>Norwegian) umlauts, Southern British /e@/ (bare, care) and French
>nasal vowels come to mind as examples. You can probably invent a
>series of complex rules to eliminate these from the inventories
>of phonemes, explain the exceptions as loanwords, analogies or
>whatever, but what would be the purpose of this?

The purpose is accuracy and thoroughness.  Insisting on
simplicity at the expense of accuracy is a particularly
pernicious type of reductionism.  Another name for it is
"oversimplification."  There are various kinds of reductionism.
A productive kind is found in breaking apart a system that is
too complex to be studied as a whole into its subsystems and
studying the subsystems independently.  This type of reductionism
is frequently applied to language because it is an extremely
complex system with many interacting levels.  Thus we have
specialists in various linguistic subsystems:  Phoneticians and
phonologists, grammarians dealing with morphology and syntax,
lexicographers dealing with the lexicon, pragmatists dealing with
matters of style and discourse, and so on.  The danger with this
kind of reductionism is that the specialists in the various
subsystems sometimes forget that their specialty is not the
entire system and that it has to be integrated with all the other
subsystems in order for the system to work.  On the whole,
however, this kind of reductionism yields good results.

Another kind of reductionism is simply refusing to allow
complexity to be considered as a possible explanation.  The
simple solution is always correct.  This kind of reductionism
elevates Occam's Razor from a heuristic to a religion.
Complicating factors are simply ignored.  As an example consider
the short term observations of the apparent motion of the sun,
moon, and fixed stars from the earth.  The simplistic explanation
of these observations is that the earth is stationary and the
heavenly bodies revolve around it.  This works fine, except for
the irregular motion of the planets.  But what would be the
purpose of working out complex explanations for the motion of the
planets that might upset the simplistic explanation of the
observations?  Let us just call the planets "wanderers" and
ignore their wayward motion, maintaining the earth as the center
of the universe.

Saying that evidence of complexity should be ignored so that we
can have simple explanations may be fine in the everyday world
where principle often gives way to expediency.  But in a
scientific context it is generally considered bad form.

<snip>

>>I just invented a phonological explanation for 'ether' - 'either'
>>so it is no longer a minimal pair (just like Comanche [papi] and
>>[pavi] aren't a minimal pair, although presumably these two words
>>are both native in Comanche).
<snip>

>To use someone else's analysis of Comanche as supporting
>evidence for your analysis of English is not very convincing.

Very perceptive.  It wasn't meant to be convincing.  On the
contrary, it was meant to show that creating silent phonemes to
block distinctions between other phonemes is not very good
methodology because it is too easy to do (see above, on the
discussion of /n/ and /ng/).  If whenever we don't want two
contrasting sounds to be phonemes we just invent a silent phoneme
so that the sounds are no longer contrasting, we no longer have a
sound basis (pun intended) for our phonological analysis.  My
point was that if the Numicists can do it for Comanche, then I
can do it for English with exactly the same result.  But I
wouldn't do it for English because, as you say, it is
unconvincing, and by implication, if it is unconvincing for
English, it is unconvincing for Comanche as well.

>If I analyzed Comanche, I would probably accept the p/v contrast
>as phonemic, even if the contrast existed only intervocally.
<snip>

Since you have already said that you consider all different
sounds that occur in a language to be phonemes, this is hardly
surprising.  But it is obvious that the Numicists feel that there
is good reason why [p] and [v] should not be considered
phonemically distinct in Comanche.  If one is not a specialist in
these languages one has to respect their view.  I only disagree
with their method of establishing this by using a silent phoneme
that is an obvious concession to the requirement that there can
be no non-phonological conditioning environments.

>It is quite common in language change for certain new phonemes
>to exist at first only in specific environments. Subsequently,
>the new phoneme is introduced into other environments by
>borrowing (from other languages or from other dialects of the
>same language) or by processes other than the one that gave rise
>to the new phoneme in the first place.

Yes, and that is how you can tell that they are being used as
phonemes.  Until that happens, though, the only reason for saying
that they are phonemes is symmetry, slot filling, or analogy.
Actual usage as a phoneme would seem to be a more viable
criterion.

>Examples:

>1. /v/, /z/, and /j/ in English. These phonemes arose from
>internal /f/, /s/ and /g > y/ in Old English under certain
>conditions (live, cheese, hedge), and they were introduced
>initially later on in loanwords (very, zero, jet) and from
>dialects (vat, vixen).

While this explanation is essentially correct for /v/ and /z/,
you seem to be somewhat confused about the development of /j/ in
native English words.  Native /g/ does shift to /y/ in English,
but it either remains in initial position (compare native 'yard'
with borrowed 'garden') or is elided intervocalically (compare
borrowed 'wagon' or 'waggon' with native 'wain', or native
'weigh' ("move"; OE 'weyan') with German '(be)wegen' ("move") and
finally (although it may be preserved graphically, it is not
preserved phonetically except as part of a diphthong; compare
'way' [wei] with German 'Weg' "way").  English /j/ developed from
a geminated g (usually <gg> in ME and <cg> in OE).  As such it
will normally correspond to <ck> in German:

   Old English    Middle English    Modern English    German

   ecg            egge              edge              Eck, Ecke
   (un)flecge     flegge            fledge            flucki (OHG)
   hecg           hegge             hedge             Hecke
   wecg           wegge             wedge             Weck
   brycg          brigge            bridge            brucka (OHG)
   hrycg          rigge             ridge             rucki (OHG)
   etc            etc               etc               etc

Since there is no initial /gg/ in English, initial
/j/ in English is usually found in loan words.  This does not
cause a problem because there is a native /j/ phoneme and the
maxim "once a phoneme, always a phoneme" allows the acceptance of
such words on a native basis and even allows new coinings with
initial /j/ like 'jingle', 'jive', and 'jitterbug'.  That's how
you can tell it's a phoneme.

>2. /b/, /d/ and /g/ in my native Hungarian. These phonemes are
>the natural development of the Proto-Finno-Ugric (PFU) internal
>clusters /-mp-/, /-nt-/ and /-nk-/, respectively. On the other
>hand, they should not exist initially, as there is no regular
>phonetic change that could produce them from PFU etyma. Yet
>Hungarian is full of words beginning with voiced stops: they are
>loanwords from Turkic, Slavic etc., and there are even a few
>words of FU origin where initial *p- and *t- changed into *b- and
>*d- ("sporadic sound change").

The dictum "once a phoneme, always a phoneme" will account for
this.

>There exist as well curious cases where an allophone acquires
>phonemic status ONLY because of the introduction of loanwords
>into a language. I am thinking of Japanese, where the phoneme /h/
>is pronounced as the voiceless bilabial  fricative (normally
>denoted as the Greek letter phi, but let's write it [ph] here)
>before the vowel  /u/ (the Hepburn transliteration is used for
>the gloss, followed by a phonemic and phonetic transcription):
>Fujimori /huzimori/ is pronounced [phujimori], fune /hune/ 'boat'
>is [phune]. Nowadays, loanwords from English are introducing [ph]
>into environments other than pre-/u/, e.g. ftku [pho:ku] 'fork',
>fairu [phairu] 'file'. According to my thinking, a new phoneme is
>being born in Japanese: the next generation of speakers will not
>necessarily know that these are "loanwords", so any phonetic rule
>based on their being loanwords will be purely ad-hoc: the only
>reason for labelling some words as loanwords will be in order to
>account for the presence of "unusual" occurrences of sounds like
>[ph] not before /u/, [sh] before /e/ etc. It is better, IMHO, to
>allow for the addition of new phonemes into the structure.

To me, the importance of this paragraph is not what it says about
the phonemicization of allophones (although the points made are
valid), but lies in the fact that it shows that you (or at least
a part of you) really do know the difference between phonemics
and phonetics.  Therefore, you should not be having this
discussion with me, but with yourself.  Specifically, the part of
you that knows the difference between phonemic /huzimori/ and
phonetic [phujimori] should try to convince the part of you that
says things like:  "the main purpose of phonemic analysis is to
provide for an unambiguous way to describe the pronunciation of
every utterance in a language" that it is wrong.  Please do this,
and I hope one of you will let me know who wins.  I certainly
hope that it is the one who knows that "/huzimori/ is pronounced
[phujimori]."

Bob Whiting
whiting at cc.helsinki.fi