Cross-linguistic phoneme correspondences
Lynne Cahill and Carole Tiberius
Information Technology Research Institute Surrey Morphology Group
University of Brighton University of Surrey
Brighton Guildford
UK UK
Lynne.Cahill@itri.brighton.ac.uk c.tiberius@surrey.ac.uk
Abstract
Cross-linguistic phoneme correspondences, or meta-
phonemes1, can be defined across languages which are
relatively closely related in exactly the same way as cor-
respondences can be defined for dialects, or accents, of a
single language (e.g. O’Connor, 1973; Fitt, 2001). In this
paper we present the theory of metaphonemes, compar-
ing them with traditional archi- and morphophonemes as
well as with similar work using “keysymbols” done for
accents of English. We describe the metaphoneme inven-
tory defined for Dutch, English and German, comparing
the results for vowels and consonants. We also describe
some of the unexpected information that arose from the
analysis of cognate forms we undertook to find the meta-
phoneme correspondences.
1 Introduction
Tiberius and Cahill (2000) presented the theory of
cross-linguistic phoneme correspondences (meta-
phonemes) with an example pilot study of the vow-
els of Dutch, English and German. The aim of this
work is to allow the type of generalisation that is
permitted by the use of phonemes with allophonic
variation to be taken one level higher, i.e. above the
level of the single language. The idea behind it is to
represent the near-identities that closely related lan-
guages such as Dutch, English and German so often
share. For example, the Dutch word ‘kat’ /kAt/ has
the English equivalent ‘cat’ /k a0 t/, and the German
‘Katze’ /kats@/2. While the consonants are largely
identical (/k/–/k/–/k/ and /t/–/t/–/ts/), the vowels are
subtly different. However, they are not distinctive
– i.e. if the / a0 / in English were replaced with /a/ it
would not sound like a different word, but rather it
would sound like a different accent. Thus our aim
is not to construct a universal phoneme set repre-
senting all phonemes occurring in a particular set
1This work was supported by ESRC grant no R000223681.
2The transcriptions are taken from CELEX (Baayen et al.,
1995) and use the SAMPA phonetic alphabet (Wells, 1989).
of languages, but we aim to capture phoneme corre-
spondences between languages such as the / a0 / – /a/ –
/A/ correspondence mentioned above. Our work is,
therefore, different from proposals put forward by
Deng (1997) who defines a set of universal phono-
logical features to be used for multilingual speech
recognition.
The three language-specific vowels discussed
above can be grouped together into the meta-
phoneme a1 a0 Aa a1 , which will be realised as an / a0 /
in English, an /A/ in Dutch, and an /a/ in Ger-
man3. Tiberius and Cahill (2000) described the
vowel metaphonemes for these three languages. In
this paper we describe a similar experiment that
looked at the consonants of the three languages. The
consonants are interestingly different from the vow-
els for a number of reasons:
a2 the consonant space is more discrete than the
vowel space, so there is less scope for small
and non-meaning-bearing distinctions within
the consonants;
a2 the phoneme inventories of the three languages
show that, while they have significantly differ-
ent vowel inventories, their consonant invento-
ries overlap greatly;
a2 while vowels were considered to occur one per
syllable (i.e. long vowels and diphthongs were
treated as single vowels), consonants can oc-
cur in clusters at either the beginning or end of
syllables;
a2 unlike vowels, consonants can be lost alto-
gether, thus leading to synchronic alternations
3We use the notation
a3 xyz a3 to denote the metaphoneme
where x, y and z are normally the sounds for the three languages
in the order English, Dutch, and German. However, this is in-
tended only as a mnemonic and does not necessarily imply that
these three sounds always occur. Metaphonemes may involve
quite complex definitions that are dependent on phonological
context as well as just the language in question.
between zero and other segments.
In this paper we present the results of our ex-
periment, first discussing the exact nature of meta-
phonemes, comparing them with archiphonemes
and morphophonemes as used in traditional ap-
proaches to morphology as well as the keysymbols
used by Fitt (2001) to define cross-accent differ-
ences. We then go on to describe the methodology
used and the results obtained. Finally we discuss
the implications of our findings both on the intended
application, i.e. multilingual lexicons, and for other
fields such as historical linguistics.
2 Metaphonemes, archiphonemes and
keysymbols
The phonemic principle, which has been with us
since the end of the nineteenth century, proposes
that sets of similar sounds, which can be distin-
guished by the phonological context in which they
occur, can be grouped together to form a single ab-
stract phoneme. The distinct sounds or phones have
been defined as allophones of the phoneme, one and
only one allophone being permitted to appear in any
particular phonological context. The metaphoneme
principle states that sets of distinct phonemes that
appear in different (but related) languages may be
grouped together in a similar way as an abstract
metaphoneme, where the conditioning factor is the
language in question rather than the phonological
context. This is the simplest case, but we also al-
low phonological conditioning to play a part in the
definition of metaphonemes. For example, we may
want to say that where English has /s/, German has
/S/ if it is in the onset and appears immediately be-
fore a /t/ but has /s/ otherwise4.
Archiphonemes (Trubetzkoy, 1939) are used to
generalise over phonemes within a language to rep-
resent cases where neutralisations arise in certain
contexts. For example, for stops that immediately
follow /s/ in English, there is no voicing distinc-
tion (‘skin’, for example, cannot be contrasted with
‘sgin’)5. Trubetzkoy proposed that in such cases we
use a different symbol to denote the underspecified
or neutralised sound. Similarly, morphophonemes
(or systematic phonemes) have been proposed by
4This is still very simplified. See below for a more detailed
discussion of this particular metaphoneme.
5In fact the realisation of such consonants is somewhere
between the voiced and voiceless forms, with minimal actual
voicing, but no aspiration that is usually associated with voice-
less stops.
generative grammarians (Chomsky, 1964) to repre-
sent situations where distinctions are neutralised in
certain morphological contexts. For example, the
voicing of the final consonant of the stem in ‘knife’
and ‘knives’ is determined entirely by the presence
or absence of the plural suffix.
Although there is a superficial similarity between
archi- and morphophonemes and metaphonemes,
there are a number of crucial differences. We should
note first that both archi- and morphophonemes
were introduced as an answer to a problem that we
do not actually face – namely the problems of vio-
lation of the phonemic principle. It is only if one
needs to insist on biuniqueness, invariance and lin-
earity that a solution to the potential problem is
needed. In the overall approach to phonology and
morphology advocated in the present work these are
simply not necessary. We allow lexical entries (or
definitions of lexical classes) to specify phonolog-
ical and morphophonological alternations without
being restricted to the phonemic principle. Thus, a
phoneme in a language can be a realisation of more
than one metaphoneme. The other most obvious
difference is that archiphonemes are defined only
within a single language, whereas metaphonemes
are defined across languages. In terms of the overall
theory of morphology, phonology and the lexicon
into which metaphonemes were designed to fit, the
generalisations represented by metaphonemes come
at a different level from archiphonemes.
The keysymbols proposed by Fitt (2001) are
much closer to our metaphonemes. The most obvi-
ous difference here is that metaphonemes range over
languages, while keysymbols are defined across dif-
ferent accents of a single language. However, this
apparently significant difference is only sustainable
if we maintain that there is a solid definition of
what is a language and what is a dialect (or ac-
cent). We would maintain that the type of lexicon
which represents related languages according to a
hierarchical definition of their similarities can be ex-
tended very simply to represent distinct dialects of
a single language in exactly the same way. How-
ever, there are practical differences in the way Fitt’s
keysymbols and our metaphonemes are employed.
Fitt assumes a text-to-speech application in which
the same words are to be pronounced, but in differ-
ent accents. We assume a more general lexicon sys-
tem, in which we may want to represent differences
in whole dialects, not just accents, so that not only
the pronunciation will be different. Fitt’s system al-
lows the definition of a single lexicon which out-
puts ambiguous strings, including keysymbols, to a
speech synthesiser which interprets the keysymbols
and disambiguates the pronunciation to get that de-
sired. In the case of metaphonemes, we anticipate
a lexical structure which allows lexical entries to be
ambiguous as to their pronunciation, but the output
of the lexicon as a whole is unambiguous, the meta-
phonemes being expanded out to their realisation in
the different languages (or dialects) as part of the
output process from the lexicon.
3 The metaphonemes of Dutch, English
and German
In order to define the metaphonemes, we con-
structed a database of around 800 cognate words
from the three languages. The database began
with orthographic forms, to which we automatically
added the phonological forms from CELEX6. We
then slightly massaged the database so that lead-
ing or trailing schwa syllables were ignored and for
most cases just the core root was left for each lan-
guage. Finally, we analysed the forms into syllabic
structures and collated the onsets, peaks and codas
for each language7.
With this information we did two things: first we
looked at the absolute correspondences, for clus-
ters and for single consonants, and their frequencies.
That is, we considered each grouping of correspon-
dences, such as:8
st+st+St
str+str+str
nd+nd+nt
m+m+m
k+k+k
This gave us both some idea of the likely corre-
spondences and some suggestions as to how phono-
logical context might affect them. We did this for
onsets only, codas only and for the two combined.
6Where there were homographs with different pronuncia-
tions, the choices between them were made manually.
7It should be noted that this process is entirely automatic,
and could be applied equally to databases of other cognate lan-
guages (e.g. French, Italian, Spanish). Indeed, it would also
be possible to construct a database that included for English
the cognates from other languages (e.g. French). There will
inevitably be gaps in the cognate mappings for any set of lan-
guages, a database that maps some English words to one lan-
guage and other words to another language would be just as
acceptable as the database we have worked with to date.
8Note that we use the ordering English, Dutch, German
throughout.
Secondly, we extracted all of the individual conso-
nant correspondences. This had to be done semi-
manually as we wanted to ensure that, in cases
such as sk+sx+S the correspondences came out
as s+s+S, and k+x+0 (e.g. ‘school’, ‘school’,
‘Schule’). From this we derived a set of tables9
which give, for each consonant in each language,
the consonants it can correspond to in the other two
languages and how often it does so in our cognate
database.
As we expected, there were many cases where the
consonants in question were almost always the same
across the languages (e.g. m+m+m). Also as we ex-
pected, the most interesting areas were where one
or more languages have different phonological con-
straints (e.g. /St/ in German onsets vs /st/ in Dutch
and English onsets) or where one or more languages
have a phoneme that the other(s) do not (e.g. /pf/ in
German, /G/ in Dutch).
3.1 Analysis of results
The tables themselves give us a great deal of in-
formation, but the whole story can only be gleaned
from both sets of data taken together. Let us now
consider in detail one small area of the analysis,
that covering the consonants /s/, /S/ and /z/. The
sounds are obviously related phonologically. /s/, /S/
and /z/ are the only sibilants that occur in all three
languages. Figure 1 shows the relevant tables for
these sounds starting from English. Just looking at
these tables tells us that for /S/ in English, there is
just a single metaphoneme worth defining, namely
a1 SsS a1 , i.e. English /S/ maps to Dutch /s/ and Ger-
man /S/. The table for /s/, however, shows us rather
more interesting things. For English /s/, Dutch has
two clear possibilities, /s/ or /z/, while German has
three, /S/, /z/ or /s/. To determine how these are re-
lated we need to look at the original correspondence
database so that we can see if there are any patterns
for the possible correspondences. The relevant en-
tries10 from the first data set for onset only are:
31 s+z+z
15 st+st+St
14 S+sx+S
6 str+str+Str
5 sw+zw+Sv
5 sp+sp+Sp
4 s+s+z
3 s+z+0
9The full tables for vowels and consonants are available at
http://www.itri.bton.ac.uk/projects/metaphon.
10This is a reduced set for simplicity.
English Dutch German
s 131 s 83 S 39
z 42 z 37
t 3 s 35
w 1 0 9
l 1 ts 5
0 1 l 2
v 1
x 1
r 1
g 1
total 131 total 131 total 131
English Dutch German
z 9 s 5 z 6
z 4 r 2
S 1
total 9 total 9 total 9
English Dutch German
S 26 s 21 S 21
k 1 0 3
l 1 k 1
z 1 z 1
0 1
d 1
total 26 total 26 total 26
Figure 1: Tables for /s/, /z/ and /S/ in English
2 st+st+0
2 spr+spr+Spr
2 sl+sl+Sl
2 sk+sx+S
for coda only they are:
21 t+t+s
16 st+st+st
11 s+s+s
5 S+s+S
4 z+s+z
3 st+st+0
3 ks+s+ks
2 0+s+s
We can see that in German, whereas /st/ appears
in coda position (corresponding strongly with /st/ in
both Dutch and English), in onset position /St/ ap-
pears corresponding with /st/ in Dutch and English.
Indeed, /s/ followed by a consonant in English and
Dutch onsets tends to correspond to /S/ followed by
that consonant in German onsets. We could spec-
ulate on many possible implications of the clus-
tering of consonants, but in the majority of cases,
the absolute correspondences across the languages
are so strong that we gain very little by consider-
ing phonological context. However, this is clearly a
case where phonological context is useful. The ta-
bles themselves suggest six possible metaphonemes
for English /s/: a1 ssS a1 , a1 ssz a1 , a1 sss a1 , a1 szS a1 , a1 szz a1 and
a1 szs a1 . The third of these we can eliminate as it is
simply the default case where all languages have the
same segment. From the data above, we can see that
the metaphoneme a1 ssS a1 is likely to be a very useful
one, as it occurs in many onset clusters.
The data above, however, allow us to say even
more. When we look at the distribution of the /s/
and /S/ in German, it is evident that a metaphoneme
that specified that English and Dutch both have /s/
in all contexts while German has /s/ in the coda and
/S/ in the onset would capture a much wider gener-
alisation, and cover 74 of the 131 English /s/ cases.
This then leaves us with the alternations that involve
/z/ in Dutch and German. We therefore propose a
metaphoneme a1 szz a1 , which is clearly evidenced by
the 31 cases of this simple correspondence for on-
sets above. However, looking more closely again,
we can see that this correspondence does not occur
at all in the coda, where English /s/ (on its own) cor-
responds to /s/ in both Dutch and German. This is
clearly a result of final consonant devoicing in these
two languages, and can be captured by making the
metaphoneme defined above phonologically condi-
tioned. Thus, English /s/ corresponds to /z/ in the
other two languages in the onset, and to /s/ in the
coda.
4 Implications of the results
The intended application of metaphonemes is hier-
archically organised multilingual lexicons that per-
mit the sharing of information at all levels (Cahill
and Gazdar, 1999), potentially useable for speech
recognition or synthesis. The use of metaphonemes
allows us to greatly increase the amount of shar-
ing of phonological information across related lan-
guages in such a multilingual lexicon. As Tiberius
and Cahill (2000) described, using metaphonemes
for the vowels alone increased the amount of phono-
logical definitions that could be shared by around
25%. While the use of consonant metaphonemes
does not lead to such significant increases in shar-
ing, we estimate that the combined figure rises to
around 40%.
Introducing metaphonemes may also be bene-
ficial with respect to the robustness of NLP sys-
tems. Knowledge about cross-linguistic common-
alities can help to provide grounds for making ‘in-
telligent guesses’ when lexical items for a particular
language are not present. For example, consider the
lexical entry for English ‘plough’. We hypothesise a
metaphoneme a1 pppf a1 (/p/ in English and Dutch, /pf/
in German) as well as a1 aUu:u: a1 (/u:/ in Dutch and
German, /aU/ in English) and a1 0xg a1 (/0/ in English,
/x/ in Dutch and /g/ in German)11. If we know that
the English word ‘plough’ has the form /plaU/ and
that the corresponding Dutch word ‘ploeg’ has the
form /plu:x/, we may predict that the German form
would be /pflu:g/. In fact, the German ‘Pflug’ has
the form /pflu:k/, due to the pervasive final conso-
nant devoicing. Thus we can see that in such a case,
metaphonemes may help us to predict a form, al-
though the result will not necessarily be fully cor-
rect. This example also illustrates the usefulness of
phonological conditioning, as we would surely want
ultimately to define all consonant correspondences
in German and Dutch to take account of the final
consonant devoicing process.
Another potential use for metaphonemes is in the
field of second language learning, where the typical
errors made by learners of a language may be deter-
mined by unconscious use of corresponding sounds
from their own language.
As well as giving a good indication of possi-
ble candidate metaphonemes, the analysis we per-
formed also gave us other information about the
three languages which is potentially of interest
to historical linguists. The analysis we did in-
volved matching the corresponding segments in
forms which are originally from identical roots.
Thus we might expect that the data can give us
clues about how the languages have changed and
diverged. For example, a zero in a possible con-
sonant position in one language suggests that that
language has lost a segment where (at least one of)
the other languages still have one. Looking at which
segments are found in such positions gives us a clue
as to which segments are most likely to be lost in
language change (at least in these languages). In-
11All of these metaphonemes are predictable from the full
correspondence tables.
deed, it transpires that the highest ranked segments
in these positions are, as one would expect, mostly
approximants, liquids and glides (/r/, /w/, /l/ etc.).
Also interesting is that of the stop consonants, the
most likely to be lost in all three languages are the
velar consonants /k/ and /g/. Another interesting re-
sult from this examination is that Dutch is appar-
ently less likely to have zeros than German, while
English is much more likely to have zeros than ei-
ther of the other two languages. (41 for Dutch com-
pared to 147 in German and 268 in English).
5 Conclusions
In this paper, we have presented the theory of meta-
phonemes. We have illustrated our theory with the
definition of cross-linguistic phoneme correspon-
dences for English, Dutch, and German. We believe
that our work has interesting benefits for speech ap-
plications. The information that can be specified in
metaphonemes can be used to tune speech applica-
tions to closely related languages in a similar way
that Fitt’s (2001) keysymbols are used to model dif-
ferent accents. The next steps in this research will
involve fully integrating the metaphonemes into a
multilingual lexicon to enable testing on a speech
synthesis system.

References

Baayen, H., R. Piepenbrock and H. van Rijn. 1995. The CELEX
Lexical Database, Release 2 (CD-ROM). Linguistic Data
Consortium, University of Pennsylvania, Philadelphia, PA.

Cahill, L. and G. Gazdar. 1999. “The PolyLex architecture:
multilingual lexicons for related languages”, In Traitement
Automatique des Langues, 40:2, pp.5-23.

Chomsky, N. 1964. Current Issues in Linguistic Theory, Mou-
ton, The Hague.

Deng, L. 1997. “Integrated-Multilingual Speech Recognition
using Universal Phonological Features in a Functional
Speech Production Model”, In Proceedings of the 1997
IEEE International Conference on Acoustics, Speech, and
Signal Processing, Vol II Speech Processing, Munich. Ger-
many. pp.1007-1010.

Fitt, S. 2001. “Morphological Approaches for an English Pro-
nunciation Lexicon.” In Proceedings of Eurospeech 2001.
Aalborg, Denmark.

O’Connor, J.D. 1973. Phonetics, Pelican Books, Great Britain.
Tiberius, C. and L.J. Cahill. 2000. “Incorporating Meta-
phonemes in a Multilingual Lexicon.” In Proceedings of the
18th International Conference on Computational Linguis-
tics (COLING 2000), Saarbruecken, Germany, pp. 1126-
1130.

Trubetzkoy, N. 1939. Grundzuge der Phonologie, Vanden-
hoeck and Ruprecht, Gottingen.

Wells, J. 1989. “Computer-coded phonemic notation of indi-
vidual languages of the European Community”, In Journal
of the International Phonetic Association, 19:1, pp.31-54.
