What is neutralization in phonology?

Answer 1

Phonological and Phonetic Representations: The

Case of Neutralization

Allard Jongman

University of Kansas

1. Introduction

The present paper focuses on the phenomenon of phonological neutralization to consider the role of phonological and phonetic representations in the production and perception of speech. Phonological neutralization involves the elimination of a phonemic distinction in a particular phonological context. Neutralization of a word-final voicing contrast is a classic example. Table 1 shows the relevant forms for German:Orthographic form

Gloss

Phonological (underlying) form

Phonetic (surface) form

Rat

advice

/Âat/

[Âat]

Rad

wheel

/Âad/

[Âat]

Table 1. Minimal wordpair with distinct phonological representations but, putatively, identical

phonetic representations.

While Rat and Rad differ in their underlying representation in terms of the voicing of the final consonant, the surface forms are identical, both ending in a voiceless alveolar stop, according to phonological analysis. In the past 20 years, phonetic research has attempted to document the extent to which this neutralization is phonetically complete. If neutralization is indeed phoneticallycomplete, the two surface forms should be phonetically identical. If neutralization is incomplete, the two surface forms

do not only differ but they should differ in predictable ways. That is, the acoustic correlates should be of the same quality as when the distinction is fully maintained.

Most phoneticresearch on neutralization has concentrated on word-final devoicing, particularly in languages such as German, Polish, and Catalan (e.g., Charles-Luce, 1985; 1993; Dinnsen and Charles- Luce, 1984; Fourakis and Iverson, 1984; Jassem and Richter, 1989; Port and Crawford, 1989; Port and O'Dell, 1985; Slowiaczek and Dinnsen, 1985).Some of these studiesreport incomplete neutralization while others document complete neutralization. Factors known to influence neutralization and therefore potentially responsible for these differences across studies include orthography (whether or not the underlying distinction is represented in the spelling, Fourakis and Iverson, 1984), speech style (careful vs. casual, Port and Crawford, 1989), and semantic expectancy (whether or not the context is predictive

of the target form, Charles-Luce, 1993).

2. Experiment 1: Manner neutralization in Korean

Since most research has concerned neutralization of a voicing distinction and results vary across studies, it is not clear a priori what one would expect to find for neutralization of a qualitatively different type. In order to address that issue, Hyunsoon Kim and I explored neutralization of manner of articulation in Korean (see Kim and Jongman, 1996, for a detailedreport). The relevant forms are shown

in (1):

 2003 Allard Jongman

Cascadilla Proceedings Project

(1) /k´t/ "to collect" /k´t + ´/ [k´.d´] (with indicative marker /´/)

/k´t + tSi/ [k´t.tS"i/ (with negative marker /tSi/)

/k´tÓ/ "outside" /k´tÓ + Èl/ [k´.tÓÈl] (with object marker /Èl/)

/k´tÓ + kwa/ [k´t.kwa] (with marker /kwa/ "and")

/k´s/ "outside" /k´s + i/ [k´.si] (with subject marker /i/)

/k´s + kwa/ [k´t.kwa] (with marker /kwa/ "and")

These three forms differ in their underlying representation. However, when followed by a suffix beginning with a consonant, all three forms surface with a plain voiceless [t]. In order to determine whether this neutralization of manner of articulation is phonetically complete, we focused on two parameters, vowel and final consonant duration. In particular, in the case of incomplete neutralization we expected to find a longer vowel duration in forms that underlyinglyend in /s/ as compared to /t, tÓ/. This

is because in Korean, like many other languages, vowels are longer when preceding fricatives relative to stops. In addition, we expected final consonants derived from underlying /tÓ/ to be longer than those derived from /t/.

Tables 2 and 3 illustrate the pattern of results that we obtained:

Underlying stem-final consonantSpeaker

/t/

/tÓ/

/s/

1

81

79

82

2

75

76

73

3

80

84

85

4

58

57

53

Table 2. Mean vowel duration (in ms) for each speaker, for minimal triplet members underlyingly ending in /t/, /tÓ/, and /s/.

Underlying stem-final consonantSpeaker

/t/

/tÓ/

/s/

1

81

79

82

2

75

76

73

3

80

84

85

4

58

57

53

Table 3. Mean consonant duration (in ms) for each speaker, for minimal triplet members underlyingly ending in /t/, /tÓ/, and /s/.

There were no differences in vowel or consonant duration as a function of the manner of articulation

of the underlying consonant. However, before claiming that these results demonstrate that neutralization

of manner in Korean is phonetically complete, it is important to realize that we measured only two acoustic cues. It is thereforepossible that differences due to underlying form would show up in some other parameter that we did not measure. Since the number of acoustic cues that would have to be measured to address this issue is potentially very large, we decided to conduct a perceptionexperiment instead to determine whether we had overlooked any cues that allowed listeners to tell the underlying forms apart.

Korean listeners were presentedwith a surface form (e.g., [k´t] derived from either /k´t/, /k´tÓ/, or

/k´s/) and asked to indicate which word they heard. Overall identification accuracy was at 32%, virtually identical to chance performance. Thus, these forms did not contain any cues that allowed listeners to decide which word they had heard. Taken together, these acoustic and perceptual results suggest that Korean manner neutralization is indeed phonetically complete. In view of some previous reports of incomplete neutralization of voicing, our finding of complete neutralization of manner raises the question whether these different results are due to the fact that neutralization of voicing and manner are qualitatively very different or to differences in the methodology employed. We therefore decided to conduct our own investigation of voicing neutralization.

3. Experiment 2: Voicing neutralization in Dutch

Similar to German, Dutch has a process of word-final voicing. The relevant forms are shown in

Table 4 (plurals are included to illustratethat the underlying consonant does surface):Orthographic form

Gloss

Phonological

(underlying) form

Plural

Phonetic

(surface) form

pond

pound

/pOnd/

/pOnd´n/

[pOnt]

pont

ferry

/pOnt/

/pOnt´n/

[pOnt]

Table 4. Minimal wordpair with distinct phonological representations but, putatively, identical

phonetic representations.

We analyzed minimal pairs produced by 15 speakers (a large corpus for this kind of neutralization study) by measuring vowel duration, consonant closure duration, and burst duration (see Warner, Jongman, Sereno, and Kemps, in press, for details). Results (averaged across words with phonemically short and long vowels) are summarized in Table 5:

Underlying word-final consonantVoiceless

Voiced

Vowel duration

148

152

Burst duration

135

129

Closure duration

78

76

Table 5. Duration (inms) as a function of the underlying voicing ofthe word-final consonant.

Statistical analysis indicated that the differences for vowel and burst duration were significant. While these differences are numerically quite small, their significance is presumably due to the large number of speakersand words employed.

We then explored whether listeners are sensitive to these differences. Instead of including all words produced by all speakers, we selected four speakers differing in the way they had produced the minimal pairs. Table 6 shows if and how each speaker used vowel and burst duration to distinguish the underlyingly voiced and voiceless tokens.Speaker

Vowel duration

Burst duration

A

++

++

B

++

+

C

0

-

D

-

-

Table 6. Extent to which each speaker used vowel duration and/or burst duration to distinguish underlyingly voiced and voiceless tokens. ++ and + indicate a relatively large or small

difference, respectively, in the expected duration , 0 indicates no difference, and - indicates a small difference in the opposite direction (see text).

For example, Speaker A had relatively large differences in theexpected direction for both vowel and burst duration. In other words, this speaker produced a substantially longer vowel and shorter burst for forms ending in an underlyingly voiced consonant. Speaker C did not differentiate forms in terms of vowel duration but produced a longer burst duration for forms ending in an underlyingly voiced consonant. In order to maximize our listeners' chances, the stimuli were blocked by speaker. The results are shown in Table 7:

Speaker

Intended final /t/

Intended final /d/

A

63*

50

B

62*

50

C

54

55

D

57

61*

Table 7. Percent 't' responses as a function of whether the speaker produced a form ending in underlying /t/ or /d/. * indicates a significant difference between the two response rates.

As can be seen from Tables 6 and 7, these results show a very good match between production and perception. Words with final /d/ and /t/ produced by speakers A and B who distinguished the voiceless andvoiced forms in terms of vowel and closure duration were correctlyidentified at significantly better than chance levels. For example, for Speaker A, there were significantly more 't' responses to words underlyingly ending in /t/ (63%) than to words underlyingly ending in /d/ (50%). Identification rates for Speaker C, who did not distinguish the differentunderlying forms in terms of vowel duration and only slightly so - but in the opposite direction - for closure duration were not significantly different. Finally, Speaker D distinguished the different underlying forms by using vowel and closure duration in the opposite direction and, lo and behold, perception exactly mimicked this pattern in that listeners perceived significantly more 't' responses when the speaker produced words ending in /d/!

We also conducted two additional perception experiments that went beyond the kind of perception experiment that is typically done in neutralization studies (as in Kim and Jongman, 1996). Both of these required the creation of duration continua by digitally editing the natural productions. In one experiment, we manipulated the cue that seemed most salient based on our findings so far.A look at Table 7 suggests that vowel duration is the most likely cue to the underlying voicing distinction.We therefore created vowel duration continuafor several words produced by each of the four speakers used previously.For example, an 11-step continuum was created from /wAt/ to /wAd/ with the original production (either

/wAt/ or /wAd/) in the middle and a vowel duration range of 50 ms. Stimuli were blocked by speaker and

continuum and listenerswere asked to respond with the word they thought the stimulus sounded more like. This is a difficult task given that vowel duration varied over only a small range (based on our production data) and the final consonant did not vary at all. Nevertheless, the results revealed that listeners were able to used vowel duration as cue. They gave significantly more 't' responses for short vowel durations as compared tolong vowel durations.

Interestingly, we obtained very similar results in the other perception experiment in which consonant closure duration continua were created. As shown in Table 5, consonant closure duration did notsystematically vary as a function of underlying consonant voicing. One may therefore expect that it does not constitute a reliable perceptual cue. However, for our continua with the original word in the middle and a closure duration range of 100 ms, listeners gave significantly more 't' responses for longer closure durations.

Taken together, the results from these two perception experiments in which acoustic parameters were systematically manipulated indicate not only that listeners can use vowel duration as a cue but they can also use closure duration as a cue even though it does not serve as a cue in natural speech. Apparently, listeners identify the only possible perceptual cue (i.e., the only property that varies within the confines of the experiment) and apply it based on its relevance in a different environment.

A number of years ago, Dan Dinnsen, one of the primary researchers in the area of phonetic and phonological neutralization, published an inventory of possible types of neutralization, taking into account both the production and perception domains. The major features of this taxonomy (Dinnsen,

1985) are reproduced here as Table 8:

Type

Production differences

Perception differences

Comments and examples

A

No

No

Standard view

Not well established

B

Yes

No

Sound change in progress

C

Yes

Yes

Non-neutralizing

German devoicing

D

No

Yes

Impossible

Table 8. Four logically possible types of neutralization when considering both production and perception (after Dinnsen, 1985).

The results of our research on Korean and Dutch indicate the need for some modifications to this table. First, Dinnsen's claim that Type A neutralization, the standard phonological view, is not well- established is challenged by our finding that neutralization of manner of articulation in Korean is complete in terms of both production and perception. Similar findings have also been obtained for the perception and production of Japanese pitch accent (Maniwa and Jongman, submitted).Types B and C are well established and presumably distinguished by the magnitude of the acoustic difference. If the difference is reliable but too small to be perceived, Type B neutralization is the result. Type C occurs when the difference is large enough to be perceived. Type D, is, of course, logically impossible. However, the results from the perception experiment on Dutch closure duration add a twist in that listeners can be made to use a cue that doesnot vary at all in production.

Finally, any discussion of phonological and phonetic representations needs to address the psychological reality of such representations. The next experiment attempted to specifically determine if listeners make use of phonological representations when perceiving speech.

4. Experiment 3: Phonological representation of [voice] in speech perception

This experiment involved the processing of Dutch verb+clitic constructions (for details, see Lahiri, Jongman, and Sereno, 1990; Jongman, Sereno, Raaijmakers, and Lahiri, 1992). In colloquial Dutch, when the clitic 'der' ("her") attaches to a preceding verb ending in an obstruent, the verb+clitic combination leads to an optional voicing alternation on the surface.Relevant forms are shown in Table

9:Verb

Underlying stem

Gloss

First ps. Sg.

Gloss

Verb+clitic

(voiceless)

Verb+clitic

(voiced)

[k{s´n]

/k{s/

'to kiss'

[Ik k{s ha®]

'I kiss her'

[Ik k{st´®]

[Ik k{zd´®]

[kiz´n]

/kiz/

'to choose'

[Ik kis ha®]

'I choose her'

[Ik kist´®]

[Ik kizd´®]

Table 9. Verb stems and voiceless and voiced verb+clitic constructions.

The verb stems for 'to kiss' and 'to choose' end in a voiceless and voiced obstruent, respectively. Crucially, for each underlying verb stem, two clitic forms can surface, one that matches and one that mismatches the underlying stem in terms of voicing.

In order to determine whether the parsing of the cliticized forms and recognition of the verbs was affected by either surface phonetic or underlying phonological representations of the verb stem, a primed auditory lexical decision paradigm was used. As illustrated in (2), listenerswould hear a target (a verb stem or infinitive form, they are identical in Dutch) preceded by a prime that was a verb+clitic construction that either matched or mismatched the target in terms of its underlying voicing representation.

(2) underlying stem-final consonant prime target

[Ik k{st´®]

/s/ [k{s]

[Ik k{zd´®]

[Ik kist´®]

/z/ [kis]

[Ik kizd´®]

Verb stems ending in /p, b/ were used in addition to stems ending in /s, z/, The same listener did not hear both the voiceless and voiced prime of the same verb. Results are shown in Figure 1.

850

800

750

/z/

/b/

700

/s/

/p/

650

600

voiced cluster voiceless cluster

Clitic Form

Figure 1. Reaction times (in ms) to voiced and voiceless clitic forms as a function of their

underlying stem-final consonants (/p, b, s, z/).

These data clearly show an asymmetry in response latencies to the same target verb depending on whether or not the listener heard the surface form that matches the phonological representation of that verb. For verbs underlyingly ending in a voiceless obstruent, responses were faster when preceded by the voiceless clitic form. Conversely, when verbs ended in voiced obstruent, responses were faster when they were preceded by the voiced clitic form. This suggests that it is not the case that both variants of a verbform are stored in the lexicon. Instead, recognition is influenced by the underlying phonological representation of the verb stem.

5. Conclusions

The Korean data suggest that neutralization of manner of articulation is phonetically complete and

as such support the standard view in phonological theory. In contrast, the Dutch data indicate that neutralization of voicing is phoneticallyincomplete. This is in agreement with some previous research

on German, although the Dutch durationaldifferences are much smaller than what has been reported for

German. One possible motivation for incomplete neutralization might be perceptual in nature. After all,

a distinction that is not completely neutralized may be easier to recover. However, that of course depends again on the magnitude or perceptual salience of the distinction and as Dinnsen (1985) already pointed out, cases have been documented in which speakers make systematic distinctions but below the level at which listeners can profit from them. In addition, the Dutch data show the reverse pattern, namely that listeners can use differences that are not reliably produced.

Assuming that both the Korean and Dutch studies have been carefully conducted, we now have solid evidence for complete neutralization in one and incomplete neutralization in the other. Both studies were quite similarin design, and both languages represent the underlying representation in their spelling,

to mention two obvious reasons for the differencesobserved. The questionthen becomes why only some

distinctions are completely neutralized. Possible answers include the following: some distinctions may

be more categorical than others. For example, the distinction between a fricative and a stop can be considered more categorical than that between a voiced and voiceless obstruent. This would account for thefinding of complete neutralization of manner but incomplete neutralization of voicing. Differences could also be due to the fact that underlying distinctions are cued by multiple acoustic parameters. In this view, degree of neutralization is determined by the extent to which each parameter can be successfully neutralized. Differencesmay also be caused by the distinction between lexical and postlexical processes,

or by paradigmatic pressure (see Mascaro, 1987).

Finally, differences may arise from the fact that phonetic implementation rules might refer to more than one representation. That is, the lexicon could have multiple underlying forms. In her insightful discussion of neutralization, Blumstein (1991) points out that if neutralization were always complete, the phonetic evidence would be consistent with the phonological analysis but would be unable to confirm it. Specifically, phonetic evidence could not rule out a lexicon with multiple underlying forms. For example, referring to Table 5, even though the singular-plural alternation is regular and predictable, the word 'pond' could have two underlying forms, /pOnt/ for the singular and /pOnd/ for the plural. However, incomplete phoneticneutralization rules out such a view. The results from the experiment with Dutch cliticized forms also suggest that a single, rather than multiple, forms are stored. In sum, then, rather than calling the concept of linguistic contrast into question, incomplete neutralization in fact supports the standard phonological analysis of neutralization.

References

Blumstein, S.E. (1991).The relation between phonetics and phonology. Phonetica 48,108-119.

Charles-Luce, J. (1985) Word-final devoicing in German: Effects of phonetic and sentential contexts, Journal of

Phonetics, 13, 309-324.

Charles-Luce, J. (1993) The effects of semantic context on voicing neutralization, Phonetica, 50, 28-43.

Charles-Luce, J. & Dinnsen,D. (1987) A reanalysis of Catalan devoicing, Journal of Phonetics, 15, 187-190. Dinnsen, D. (1985) A re-examination of phonological neutralization, Journal of Linguistics, 21, 265-279.

Dinnsen, D. & Charles-Luce, J. (1984) Phonological neutralization, phonetic implementation and individual differences, Journal of Phonetics, 12, 49-60.

Fourakis, M. & Iverson, G. (1984) On the 'incomplete neutralization' of German final obstruents, Phonetica,41,

140-149.

Jassem, W. & Richter, L. (1989) Neutralization of voicingin Polish obstruents, Journal of Phonetics, 17, 317-325. Jongman, A., Sereno, J., Raaijmakers, M., and Lahiri, A. (1992). The phonological representation of [voice] in

speech perception. Language and Speech 35,137-152.

Kim, H., and Jongman, A. (1996). Acoustic and perceptual evidence for complete neutralization of manner of articulation in Korean.Journal of Phonetics 24, 295-312.

Lahiri, A., Jongman, A., and Sereno, J. (1990). The pronominal clitic [d´r] in Dutch: A theoretical and experimental

approach. In G.E. Booij and J. van Marle (eds.), Yearbook of Morphology, Vol. 3(pp.115-127). Dordrecht: Foris Publications.

Mascaro, J. (1987). Underlying voicing recoverability of finally devoiced obstruents in Catalan. Journal of

Phonetics 15,183-186.

Port, R. & Crawford, P. (1989) Incomplete neutralization and pragmatics in German, Journalof Phonetics, 17, 257-

282.

Port, R. & O'Dell, M. (1985) Neutralization of syllable-final voicing in German, Journal of Phonetics,13, 455-471. Slowiaczek, L. & Dinnsen, D. (1985) On the neutralizing status of Polish word-final devoicing, Journal of

Phonetics, 13, 325-341.

Warner,N., Jongman, A., Sereno, J., and Kemps, R. (in press). Incompleteneutralization and other sub-phonemic durational differences in production and perception: evidence from Dutch.Journal of Phonetics.

website: neutralaccent.com

Answer 2

Neutralization in phonology occurs when two distinct phonemes are pronounced the same in a specific phonological context. This can happen in certain environments where the contrast between two phonemes is lost. Neutralization is often a result of phonological processes such as assimilation or reduction.