NEUTRALIZATION.?!

The phonetics – phonology issue in the analysis

of word-final obstruent voicing

Klaus J. Kohler
University of Kiel
Kiel, Germany
kjk@ipds. uni-kiel.de

ABSTRACT
In German and Polish, the voicing contrast in obstruents is excluded from
word-final position. Prague Circle phonology treated this as phonological
neutralization with phonetic voiceless exponency. Generative Phonology
created underlying phonological contrasts, which experimental
phoneticians started testing for phonetic substantiation at the surface
level. Laboratory Phonology has adopted this paradigm and continues to
maintain that neutralization is incomplete in spite of differences being
too small for communicative relevance. The debate can be terminated by
substituting a polysystemic syntagmatic approach for the monosystemic
paradigmatic one.

1. From the Archiphoneme to Incomplete Neutralization

In a fair number of languages, the voicing distinction in obstruents is limited
to certain positions. In Polish, the domain of these limitations is the phrase,
across morpheme and word boundaries: all final obstruents are voiceless
before pause, and internal obstruent clusters are either all voiceless or all
voiced depending on the voicing feature of the last obstruent [1]. In German,
the domain is the word: all obstruents are voiceless in final position (also
morpheme-final before consonant, e.g. möglich /k/ vs. hüg(e)lig /g/) [2].
These phonotactic restrictions were conceptualized in Prague Circle
phonology as neutralization of a phonological opposition to an archi-
phoneme, which incorporates all the features that the elements of the
neutralized pair have in common, and thus does not coincide with either of
them, but which is represented phonetically by one or the other [3]. In the
cases quoted, it is the voiceless obstruent in German, the voiceless or the
voiced one in Polish.
For alternations between archiphoneme and voiced or voiceless
obstruent in morphological paradigms, as between German Bund [t] –
Bunde, Bünde [d], bunt [t] – bunte [t], Prague Circle linguistics, just like
American Structuralism, set up a morphophonological level, separate from
phonology in the narrow sense. This conception of levels of linguistic
description changed with the advent of Generative Grammar, which denied
systemic phonology an independent status and integrated phonological
correspondences in morphological paradigms in the phonological component
of a grammar. Phonological alternations were represented by the same
contrast of underlying distinctive feature bundles across all positions, from
which the neutralized form was derived by phonetic rule, e.g. word-final
devocing of /d/ in Bund [t]. Thus neutralization was treated as a phonetic
process operating on a phonological distinction.
So, when researchers started taking these new phonological solutions
into the laboratory, it was a natural consequence for them to ask whether the
postulated non-neutralization in phonology was also supported by a phonetic
contrast, which had not been detected previously due to missing
measurements. The first to raise this incomplete neutralization issue for
German were Mitleb [4], Port et al. [5], O'Dell and Port [6], followed by
more extensive accounts and discussions in Port and O'Dell [7]. These
researchers stood in the Haskins Labs tradition of analysing the
voiced/voiceless contrast in English word-final plosives with reference to the
durations of the preconsonantal vowel, the closure and closure voicing, and
the release, which they applied to German words. Consistent differences
were found, especially in the duration of the vowel, in citation form
pronunciation. They were small (15 ms of vowel and burst durations, 5 ms of
closure voicing on average), but in the same direction as in the non-
neutralizing word-medial context, i.e., longer vowel, shorter burst, and more
closure voicing in the underlying voiced plosive, some of them reaching
statistical significance. The production data were supplemented by listener
judgments, which identified the produced member of a pair poorly, but
significantly better than chance (59% overall). The authors concluded from
these results, supported by inferential statistics, that German had incomplete
neutralization of the underlying word-final voicing opposition in obstruents,
and that these phonetic differences would have to be taken into account in
the phonology of German by the introduction of implementation rules of the
generally maintained phonological contrast.
Fourakis and Iverson [8] argued convincingly, on the basis of data from
orthographic vs. free elicitation of relevant words, that the phonetic
differences were caused by hypercorrection of isolated words in a reading
task, and that there were no grounds for postulating incomplete
neutralization in the phonology of German. They also pointed out that weg
(although etymologically related to the noun paradigm Weg : Wege) was not
part of a morphological alternation at all, now being an adverb with a short
vowel. Another crucial point in the evaluation of these data is that the
authors found it necessary to explain the meanings of some of their words to
their subjects before the actual tests, viz. Alb 'elf' vs. Alp 'mountain pasture',
schrag (imperative form of obsolete schragen) 'join wooden poles cross-wise
or slanting' vs. schrak (past tense of intransitive schrecken, which only
occurs in erschrecken) 'had a fright', Weck 'breakfast roll' (southern dialect
word). In view of the complete lack of context for words, which in addition
are not all part of the subjects' common vocabulary or are unusual
morphological forms, the test items assumed the status of logatomes.
Furthermore, the data were collected in the US from German speakers who
had spent considerable lengths of time in an English-speaking environment
and had no doubt been made aware of the need to differentiate final
orthographic <b>, <d>, <g> from <p>, < t>, <k>. In view of these flaws in
experimental design, these data were not able to substantiate the claim of
incomplete neutralization. But the discussion has continued ever since.
Slowiaczek and Dinnsen [9] applied the same phonological paradigm
with the same experimental design flaws to neutralization in Polish, and
obtained the same type of small differences. On the basis of such phonetic
measurements in underlying phonological distinctions, Dinnsen [10,11]
finally concluded that careful phonetic studies would reveal that all declared
neutralization rules in the phonologies of languages were in effect not
neutralizing at all. Not only did he accept the validity of the phonetic data
without query, he also put forward theoretical phonological arguments to
counteract the objection that the acoustic production differences are too
small to be perceived. As in generative phonology grammars are assumed to
be abstract representations, neutral with respect to speaker and hearer, the
existence of distinctions in production but of neutralization in perception
would entail a different theoretical conceptualization. The speaker-hearer
indifference would vanish, neutralization would become part of phonological
perception, and production differences would be linguistically irrelevant.
This would have “awkward consequences” for allophonic rules, which
“describe production differences [that] are not generally self-discriminable
by native listeners” So, if “phonological rules specify segments differently
only if their production differences are perceptually salient, phonological
descriptions would exclude all allophonic statements” [11, pp. 271f].
There are two flaws in this line of argument. First of all, it mixes up
non-discriminability on the basis of an acoustic distinction being below the
Just Noticeable Difference [JND] of the perception mechanism, and non-
discriminability of contextual variation in category allocation experiments.
Naïve listeners in most English varieties may not be able to “hear” the
difference they make between clear and dark /l/ initially and finally in
English little, when asked about the presence of a sound distinction, but they
will immediately recognise the absence of the right /l/ in final position and
categorize it as ‘foreign’ or ‘Irish’. The lack of perceptual relevance of
acoustic measurements in the incomplete neutralization issue, however, has
to do with the JND threshold. The second flaw concerns the conception of a
language as an abstract system indifferent as to speaker and hearer, and thus
dissociated from speech communication, but nevertheless buttressed with
reference to speaker data obtained in problematic experimental lab speech
designs.

2. Refuting Incomplete Neutralization in Polish

Jassem and Richter [1] replied to [9] with an experimental analysis that
removed the flaws of previous data acquisition and placed phonological
description in a framework of speech communication rather than on a level
of abstract linguistics. The findings of [9] were reexamined with materials
that were to meet the following conditions:
• speakers are recruited within their native-language country
• speakers do not see the materials in orthographic form
• materials are not of a metalinguistic type but consist of unprepared
utterances elicited in a natural dialogue situation.
• speakers represent a homogeneous group with regard to dialectal
variation.
The results show a slight tendency of mean durations of vowels towards
greater, and of consonants and consonant releases towards smaller, values for
+voice. The differences are extremely small and not always regular (4 ms for
vowel and for consonant durations), and none of the differences reach
statistical significance. In a subsequent listening test, subjects indicated a
strong bias for phonetic voicelessness corresponding to underlying –voice as
well as +voice in final neutralizing position. So the conclusion is that “data
obtained from materials recorded under conditions approaching a
spontaneous discourse situation give no ground for rejecting the traditional
view that in positions of phrase-final devoicing and internal assimilation,
neutralization of the voice feature in Polish obstruents is phonetically (and
eo ipso phonologically) complete. … Extrapolating from experiments
performed earlier elsewhere under conditions of lesser naturalness than ours,
as well as from our own results, there appears good reason to conclude, as a
generalization, that if any linguistic-phonetic (phonological) descriptions are
to be tested experimentally, it is mandatory to ensure maximum naturalness
of the materials.” [1, p.124]

3. The continued issue of Incomplete Neutralization in German

In spite of the critical assessment of the research into incomplete
neutralization of voice in final obstruents (see also [12] for a critique of the
German data), researchers persisted in pursuing the question, with little or no
regard to Wiktor Jassem’s mandatory guidelines for data acquisition. Latterly
the Laboratory Phonology framework became the ideal environment to carry
phonological solutions of final voicing restrictions into the laboratory for
phonetic substantiation and thus kindle renewed investigation into
incomplete neutralization. A recent paper [13] is part of this paradigm.
In a production experiment, subjects successively heard 48 randomized
non-words (disyllabic minimal pairs with intervocalic /bp/ (8), /dt/ (7) , /gk/
(9) of German phonotactic structure), each embedded in a meaningful
German sentence, pronounced by a male speaker, in which the non-word was
decodable as a masculine plural noun. They were asked to respond to each
heard stimulus sentence by a given sentence frame in which they were to
place the noun in the singular, thus moving the plosives into final position.
16 native speakers of German from the Cologne area, 9 women, 7 men,
participated in the experiment. In the disyllabic prompts, the vowel was on
average 28 ms shorter before voiceless stops, closure duration and VOT
were on average 21 ms and 41 ms longer, all statistically highly significant.
So, the acoustic differences in the disyllabic pairs were strongly marked. In
the monosyllabic responses, however, only vowel duration reached signific-
ance, and on average was only 8 ms shorter before the voiceless plosive.
This minimal value is below the JND, and there are thus worlds between the
signalling of the underlying phonological contrasts in the two contexts.
The authors maintain that subjects could not be influenced by
orthography in the production experiment since the task did not include
written words but only aural stimuli which had to be output in modified
form. However, since the subjects were no doubt undergraduates as well as
PhDs from the Cologne Linguistics Department, a fairly high degree of
literacy can be assumed, even to the extent of a conscious awareness of the
relationship of sounds and letters. Furthermore, when naive language users
are to repeat real spoken words they recognise them as such and are then
able to output them again. If they also have to modify them this already
introduces an interfering variable. But if they cannot categorize the auditory
stimuli linguistically as existing words they will find it very difficult to even
just repeat them. Categorization via an orthographic representation will help
them, but this instrument is only available to literate people and works the
more efficiently the greater the degree of literacy and awareness of the sound
– letter relationship. The fact that a considerable 12% of the data (16 * 48 =
768 items), i.e. an average of 5-6 items per speaker, had to be excluded
because they were either incorrectly remembered or reproduced with a lot of
hesitation, points in the same direction. So, orthographic influence cannot be
confidently excluded.
In the subsequent perception experiment, 8 listeners had to judge 192
sentence items from the production experiment by referring to the
orthographic alternatives presented for each word (e.g. <Drud> vs. <Drut>).
They were instructed that exactly half of the stimuli were from the set <b, d,
g> and half were from the set <p, t, k>. This introduced a bias towards
voicing in the experimental design despite the absence of strong cues for it in
the production data, and also raised the number of false alarms in voiceless
items. The overall average accuracy of voiced or voiceless identification was
54%, pointing to high proportions of misses and false alarms. Participants
ranged from 50% to 59%. This is rather poor performance, compared with
98% correct identification of the prompts. In view of these results, subjects’
judgments cannot be based on voiced/voiceless cue perception but are
random responses, in spite of statistical significance of the small vowel
duration difference. This means that the existence of incomplete
neutralization in German has not been proved in this renewed perceptual
investigation either.
It is thus unjustified to conclude “…our results demonstrate that IN
effects in perception are robust to the extent that they occur in a forced-
choice paradigm even with a multitude of voices ... These experiments
therefore address important concerns with previous investigations of incom-
plete neutralization. Our results indicate that IN is a robust phenomenon
whose phonological implications should be taken seriously.” [13, p. 1725]

4. A different approach to Neutralization

In this Laboratory Phonology approach, phonetic data are no longer analysed
and interpreted as sound patterns in speech communication but are related to
abstract phonological categories which are removed from speech and have
their aim in themselves, and inferential statistical tests are applied to validate
them. This is what the investigation into incomplete neutralizations is all
about. And since their acoustic manifestations and perceptual consequences
are so weak and unclear, and based on flawed experimental designs far
removed from natural speech communication, they cannot be postulated as
having any relevance in human interaction by sound.
It is high time that researchers in speech science heard Wiktor Jassem’s
lone cry from the phonetic wilderness, and stopped wasting enormous
resources of time, manpower and public money chasing a phonology-
induced phantom, and that they turned their scientific skills instead to the
elucidation of more interesting questions of real speech communication. A
look at the problems German children have learning to write can convince
anybody, without sophisticated speech signal analysis, that, e.g., Kind sounds
like Kint to them, because that is how they spell it before they learn the
morphological rules.
Neutralization of the voicing feature in word-final obstruents in
Standard German and its associated dialects is a phonetic reality in speech
interaction, just as it is in Polish and no doubt other languages, whereas
incomplete neutralization is an experimental artifact in the wake of
generative phonology. But in colloquial German the phonetic reality goes
even beyond the word and is becoming a feature of the phrase, operating
across word boundaries, as was described mutatis mutandis for Polish by
Wiktor Jassem. In formally and semantically cohesive lexicalized phrases,
such as außer Rand und Band, auf Schritt und Tritt, Lug und Trug, Dick und
Doof, lieb und teuer, ab und zu, the realization of the final plosive in the first
word fluctuates along a VOT scale from voiceless aspirated to unaspirated to
voicing, irrespective of the phonological voicing feature of the plosive in the
morphological paradigm, only depending on the syntagmatic cohesion of the
utterance in spontaneous communication. If the speaker breaks the
articulatory flow after the first word by sentence-prosodic means – a tonal
boundary, articulatory deceleration, a pause, glottal onset of the following
vowel – the realization is voiceless aspirated. The greater the syntagmatic
prosodic cohesion between the words in the phrase, the greater the likelihood
that the plosives lose their aspiration and become phonetically voiced, no
matter whether they are “phonologically” /bdg/, as in Rand vs. Ränder, Zug
vs. Züge, and lieb vs. liebe, or /ptk/, as in Schritt vs. Schritte, dick vs. dicke,
and ab, which is outside a paradigm, and whose final plosive must today be
allocated to /p/, in spite of OHG aba, MHG ab(e).
Figure 1 gives some examples produced with varying degrees of
cohesion in a pilot production test, where subjects had to read a list of
sentences that contained such lexicalized phrases, in which the first word
was replaced by ___, to detract from orthographic representation of the
target word, e.g. Er gerät nicht selten außer ___ und Band. Er folgt ihr auf
___ und Tritt. Wir erledigen das Geschäft ___ um Zug. Laurel und Hardy
sind bei uns als ___ und Doof bekannt. Sein Job ist ihm ___ und teuer. Ich
schlage vor, dass wir uns ___ und zu in der Mensa treffen. Subjects had to
repeat each sentence several times.

Figure 1: Spectrogram and phonetic segmentation of the German

phrases a “Rand und Band”, b “Schritt und Tritt” (female speaker),
and c “Zug um Zug”, d “Dick und Doof” (male speaker).
 These syntagmatic processes spanning the exponents of a paradigmatic
voicing contrast can be related to a rhythmical principle that is operative in
such fixed dual-element phrases (cf. [14]). In Germanic languages, their
rhythmicity is governed by the number of syllables (in English Stan and
Ollie is more rhythmical than the reverse), or by high front vs. back/low, and
short vs. long vowel (the German equivalent Dick und Doof is more
rhythmical than the reverse, and reverses the order of physical appearance in
Laurel and Hardy). Friction noise in the first part of such phrases interrupts
the rhythmic flow. That is an additional reason why Doof is not in first
position, and why Jane Austen’s Sense and Sensibility was translated as
Gefühl und Verstand, reversing the semantic order of the English title.
So, the voiced/voiceless contrast is not only phonetically absent from
final obstruents in German words, plosives being realized as voiceless
aspirated in citation form pronunciation, but there are also
unaspirated/voiced realizations in fixed syntagmatic frames with high
cohesion where rhythmicity becomes the binding principle. This
syntagmatic domain of voicing exponency of German word-final
obstruents as part of speech rhythm needs systematic investigation in
production and perception, and will not only advance our knowledge of
segmental structures in phrase phonology but will also give us new insights
into what constitutes speech rhythm.
Of course, incomplete neutralization would never have become an issue
of phonetic analysis if American Structuralism had applied a polysystemic
approach to phonology, as did Firth and his School [15, 16], instead of
postulating an all-encompassing monolithic phoneme system, because then
there would not have been anything to neutralize. The archiphoneme of the
Prague Circle was already a step in this direction, and the extension of the
issue of voicing exponency to the syntagmatic domain fits nicely into this
polysystemic approach, providing another case that demonstrates the need to
bridge the segment – prosody divide [17].

5. Epilogue
The different ways neutralization has been handled since the days of the
Prague Circle may be given melodic expression.
a Trubetzkoy made a phonological statement
There’s neutrali'zation.
shown in Figure 2a: low pre-head, followed by an f0 rise-fall spread
over the stressed and the following unstressed syllable, masked
during the voiceless fricative;
b Haskins phoneticians queried its phonetic adequacy
 Neutrali'zation?
shown in Figure 2b: low pre-head, followed by a high f0 rise from
the stressed-syllable onset to the end of the following unstressed
syllable, masked during the voiceless fricative;
c Dinnsen and the Laboratory Phonology community rejected it with
phonological incredulity and indignation “Lady Bracknell” style
[18]
Neutrali'zation?!
shown in Figure 2c: high pre-head, followed by a late high-rising
valley contour with breathy intensification of the stressed syllable,
masking f0 up to the sonorant of the unstressed syllable;
d Wiktor Jassem brought it back to the phonetic level of American
Structuralism with politely contrastive but compelling reference to
phonetic reality
There’s neutrali'zation!
shown in Figure 2d: low pre-head, followed by an f0 rise-fall-rise on
the stressed and the following unstressed syllable, the fall being
masked by the voiceless fricative;
e I move the discussion from the paradigmatic monosystemic to the
syntagmatic polysystemic level in a final conclusion
There’s 'nothing to 'neutralize.
shown in Figure 2e: low pre-head, followed by an f0 rise in the first
stress foot 'nothing to and a subsequent early low f0 fall in the
second stress foot 'neutralize.
The prosodic embedding of segmental neutralization in this paper is a
small personal tribute to Wiktor Jassem on the occasion of his 90th birthday
in appreciation of his broad spectrum of contributions to phonetic science
from segments to rhythm and intonation.
 Figure 2: Spectrograms, f0 curves and syllabic segmentation of a
“There’s neutralization.” (statement), b “Neutralization?” (confirmation
question), c “Neutralization?!” (emphatic contradictory question), d
“There’s neutralization!” (contrastive statement), e “There’s nothing to
neutralize.” (matter-of fact conclusive statement). Speaker: the author.

REFERENCES
[1] Jassem, W. and Richter, L. 1989. Neutralization of voicing in Polish
obstruents. Journal of Phonetics 17, 317-325.
[2] Kohler, K. J. 1995. Einführung in die Phonetik des Deutschen. Berlin:
Erich Schmidt Verlag (2nd ed.).
[3] Trubetzkoy, N. S. (1939). Grundzüge der Phonologie. (3rd ed. 1962
Göttingen: Vandenhoeck & Ruprecht).
[4] Mitleb, F. 1981. Segmental and Non-segmental Structure in Phonetics:
Evidence from Foreign Accent. PhD diss. Indiana University,
Bloomington.
[5] Port, R., Mitleb, F., and O'Dell, M. 1981. Neutralization of obstruent
voicing in German is incomplete. Journal of the Acoustical Society of
America 70, S13, F10.
[6] O'Dell, M., and Port, R. 1983. Discrimination of word-final voicing in
German. Journal of the Acoustical Society of America 73, S31, N17.
[7] Port, R., and O'Dell, M. 1985. Neutralization of syllable-final voicing in
German. Journal of Phonetics 13, 455-471.
[8] Fourakis, M., and Iverson, G. K. 1984. On the 'incomplete neutralization'
of German final obstruents. Phonetica 41, 140-149.
[9] Slowiaczek, L. M. and Dinnsen, D. A. 1985. On the neutralizing status of
Polish word-final devoicing. Journal of Phonetics 13, 325-341.
[10] Dinnsen, D. A. 1983. On the Characterization of Phonological
Neutralization. Bloomington: Indiana University Linguistics Club.
[11] Dinnsen, D. A. 1985. Re-Examination of Phonological Neutralization.
Journal of Linguistics 21, 265-279.
[12] Kohler, K.J. 2007. Beyond laboratory phonology: The phonetics of
speech communication. [In:] M. J. Solé, P. S. Beddor, Ohala, M. (Eds.)
Experimental approaches to phonology, 51-53. Oxford: Oxford
University Press.
[13] Röttger, T. B., Winter, B., and Grawunder, S. 2011. The robustness of
incomplete neutralization in German. Proc. 17th ICPhS, Hong Kong,
1722-1725.
[14] Kohler, K.J. 2009. Rhythm in speech and language. A new research
paradigm. Phonetica 66, 20-45.
[15] Firth, J.R. 1948. Sounds and prosodies. Trans. Philological Soc. 1948:
127– 152.
[16] Hawkins, S.; Smith, R. 2001. Polysp: a polysystemic, phonetically- rich
approach to speech understanding. Ital. J. Ling. 13: 99– 188.
[17] K. Kohler, K. 2011. Communicative functions integrate segments in
prosodies and prosodies in segments. Phonetica 68, 26-56.
[18] Kohler, K.J. 2011. On the interdependence of sounds and prosodies in
communicative functions. Proc. 17th ICPhS, Hong Kong, 19-27.