voicing variability and formal phonology - stanford...

Incomplete final devoicing Determining underlying voicing

Voicing Variability and Formal Phonology

Marc van OostendorpMeertens Instituut & Leiden University

Variation, Gradience and Frequency in Phonology

09.08.07

Voicing Variability and Formal Phonology Variation, Gradience and Frequency in Phonology



I Final devoicing seems to pose two major problems for acategorical theory of phonology:

I The output is gradientI The input is predictable on the basis of corpus distribution

I From this, people have drawn the conclusion that either thedata are wrong or non-phonological, or that formalphonology is wrong

I We offer instead a refined, but classical phonologicalaccount of these experimental data



The issue


Incomplete final devoicingThe issueFaithfulnessTurbidity Theory

Determining underlying voicingThe issueVoicing in DutchConclusions



The issue

Final Devoicing

I Catalan:I gris ‘grey (M)’ - griz@ ‘grey (F)’I gos ‘dog (M)’ - gos@ ‘dog (F)’

I Dutch:I kwaa[t] ‘angry (PRED.)’ - kwad@ ‘angry (ATT)’I laat ‘late (PRED.)’ - lat@ ‘late (ATT)’

I Polish:I klup ‘club’ - klubi ‘clubs’I trup ‘corpse’ - trupi ‘corpses’



The issue

Phonetic Timing



The issue

Listening task

I if asked to randomly guess whether a given instance [bunt]corresponds to (German) /bund/ ‘league’ or /bunt/‘colourful’, speakers will guess correctly (60 to 70 per cent)

I “If [these words] were the same, then in a listening taskyou would expect 50 percent correct (pure guessing — likeEnglish too and two would show). If contrastive, one wouldexpect at least 99 percent correct identification under goodlistening conditions with motivated subjects (just like Bundeand bunte would show).” (Port & Leary 2005)



The issue

Possible responses

1. The data are wrong, for instance because they have beenacquired under suspicious laboratory conditions;

but theyhave now been replicated for many languages, and weneed to account for laboratory behaviour as well

2. Phonology is wrong; but this throws away decades of solidresults

3. This is all phonetics; but that means direct access ofphonetics to the lexicon

4. We have to integrate these facts into a classical model



The issue

Possible responses

1. The data are wrong, for instance because they have beenacquired under suspicious laboratory conditions; but theyhave now been replicated for many languages, and weneed to account for laboratory behaviour as well






The issue

Possible responses


2. Phonology is wrong;

but this throws away decades of solidresults





The issue

Possible responses







The issue

Possible responses



3. This is all phonetics;

but that means direct access ofphonetics to the lexicon




The issue

Possible responses







The issue

Final devoicing is phonological

I ik heb ‘I have’ [Ik hEp]I hebben ‘to have’ [hE.b@n]I ik heb ’m ‘I have him’ [Ik hE.p@m]

(Booij 1995)



The issue

Formal analyses

I There is a phonetic paradigmatic effect; ‘Word-basedphonetics’ (Pierrehumbert 2002) is a possibleimplementation of this.

I The laryngeal contrast between voiced and voicelessobstruents is ‘enhanced’ by other features (Avery and Rice1989).

I Both of these complicate the relationship betweenphonology and phonetics



Faithfulness






Faithfulness

Containment and Correspondence

1. Correspondence Theory: There are separate input andoutput representations, as well as correspondenceconstraints between elements of these (McCarthy andPrince 1995)

2. Containment Theory: The input is contained in the output,therefore all faithfulness constraints can be read off thesurface representation (Prince and Smolensky 1993, VanOostendorp 2005).



Faithfulness

Correspondence

k

k

l u

u

k

k u

input

output Universe { k1, l2 ,u3 ,k4 } ∪ { ka, ub,kc ,ud }Relations C (k1,ka)∧C (u3 ,ub)∧C (k4,kc)



Faithfulness

Containment

I Containment. Every element of the phonological inputrepresentation is contained in the output. (There is nodeletion.)



Faithfulness

Containment: Prince and Smolensky 1993

I PARSE: All elements should be ‘parsed’ in the phonologicalstructure (no deletion.)

I FILL: Do not allow empty elements. (No insertion.)



Faithfulness

Containment Representation

Φ

k l u k ∅ Universe { k1, l2 ,u3 ,k4,∅5 }Relations Dφ(Φ,k1)∧Dφ(Φ,u3)∧Dφ(Φ,kk )∧Dφ(Φ,∅5)



Faithfulness

Occam’s Razor and Containment

I PARSE-C: Every consonant needs to be affiliated toprosodic structure

I FILL-V: (Nuclear) syllable slots need features.



Faithfulness

Problems with the Prince & Smolensky Interpretation

I features should also not be allowed to ever spread to anepenthetic vowel

I how do we prevent spreading from happening everywherein every language?



Faithfulness

Consistency of Exponence

I “No changes in the exponence of aphonologically-specified morpheme are permitted.”(McCarthy and Prince 1993, 1994)



Faithfulness

Explanation

“[Consistency of Exponence] means that the lexicalspecifications of a morpheme (segments, prosody, or whatever)can never be affected by Gen. In particular, epentheticelements posited by Gen will have no morphological affiliation,even when they lie within or between strings with morphemicidentity. Similarly, underparsing of segments — failure to endowthem with syllable structure — will not change the make-up of amorpheme, though it will surely change how that morpheme isrealized phonetically. Thus, any given morpheme’sphonological exponents must be identical in underlying andsurface form.”



Faithfulness

CoE Representation

Φ

k l u k u

M Universe { k1, l2 ,u3 ,k4,u5 }Relations DM (M,k1)∧DM (M,l2)∧DM (M,u3)∧DM (M,k4)

Dφ(Φ,k1)∧Dφ(Φ,u3)∧Dφ(Φ,kk )∧Dφ(Φ,u5)



Faithfulness

Faithfulness constraints (coloured versions)

I PARSE-φ(x): The morphological element x must beincorporated into the phonological structure. (No deletion.)

I PARSE-µ(x): The phonological element x must beincorporated into the morphological structure. (Noinsertion.)



Turbidity Theory






Turbidity Theory

The trouble with features

input outputx x

F

x x

F



Turbidity Theory

The trouble with PARSE-F

input outputx x

F

x x

F



Turbidity Theory

Two relations instead of one

I projection: an abstract, structural relationship holdingbetween a segment and the feature (roughly equivalent tonotions of ‘Licensing’).

I pronunciation: an output relationship that holds betweenthe feature and the segment and describes the outputrealization of structure.



Turbidity Theory

Turbidity in Goldrick’s work

I /ka+tiko/ → [katiko] ‘mushroom’

I /ka+oto/ → [ko:to] ‘fireplace (DIM)’I /ka+ezi/ → [ke:zi] ‘moon (DIM)’

(Luganda)



Turbidity Theory

Turbid representation

µ µ

6@@R?6

a oTurbidity presupposes Containment.



Turbidity Theory

Stray Erasure (Turbid version)

I The phonetics only interprets features that stand in apronunciation relation to a segment in the phonology.



Turbidity Theory

Turbidity as Containment (Revithiadou 2006)

I we take projection lines to represent the lexical state ofaffairs, that is, to be part of the lexical representation of amorpheme [. . . ]. In conformity with [Consistency ofExponence], therefore, they cannot be altered by Gen.



Turbidity Theory

Reciprocity

I RECIPROCITYVF (RV

F ): If a vowel V entertains aprojection relation with a feature F, then F must entertain apronunciation relation with the vowel V.

I

x6

F



Turbidity Theory

FinDev

I [voice] cannot entertain a pronunciation relation with anobstruent in the coda.



Turbidity Theory

Candidates

a. [kwa:d˚

] b. [kwa:d]

k w a: d6

voice

k w a: d6?voice



Turbidity Theory

Miniature TypologyI FINDEV�RECIPROCITYV

F : Final Devoicing (Catalan,Dutch, etc.)

/kwa:d/ FINDEV RECIPROCITYVF

kwa:d *!+kwa:d

˚*

I RECIPROCITYVF�FINDEV: Final Devoicing (Spanish,

English, etc.)

/kwa:d/ RECIPROCITYVF FINDEV

+kwa:d *kwa:d

˚*!



The issue






The issue

(Weak) past tense formation

I -/t@/ after roots ending in underlyingly voiceless obstruents:kook-te ‘cooked’, raap-te ‘gathered’, praat-te ‘talked’

I -/d@/ after all other stems: leev-de ‘lived’, meld-de‘mentioned’, ren-de ‘ran’

Past-tense formation thus reveals the underlying voicing ofobstruents.



The issue

What happens with new (nonsense) verbs

I First person singular does not have any overt suffix, hencefinal devoicing (ik leef ‘I live’, ik kook ‘I cook’, ik kam ‘I live’,ik ren ‘I run’)

I This allows for a straightforward Wug test:I ‘What is the past tense of ik le[p], ik sta[x], ik draa[s]?’



The issue

Baayen and Ernestus (2003)

Responses ending in -de and -te, by type of stem-finalobstruent (CV:C words)

TYPE de teP 3 97T 4 96S 55 45F 64 36X 81 13



The issue

Correspondence to corpus data



The issue

The puzzle

I From these facts, Baayen and Ernestus (2003) concludethat speakers have knowledge of the statistical distributionin the corpus

I However, it is unclear that we can establish a direct causalrelation here

I and it is left unexplained why the statistical distribution isthe way it is



The issue

The hierarchy and its surprises

I P > T > F > S> XI “[T]he type of the final obstruent itself is an important

predictor of voicing. Surprisingly, a mirror image of thehierarchy of phonological strength emerges from the data[. . . ] Dutch shows a preference for words to use finalobstruents with a high cue validity. More research is clearlyrequired here.” (B&E 2003:30)



The issue

Possible responses

I The data are wrong

but again there is no indication for thisI The data are non-phonological but there is a clear

interaction with ‘real’ phonologyI Phonology is wrong but this throws away decades of solid

resultsI We have to integrate these facts into a classical model



The issue

Possible responses

I The data are wrong but again there is no indication for this

I The data are non-phonological but there is a clearinteraction with ‘real’ phonology

I Phonology is wrong but this throws away decades of solidresults

I We have to integrate these facts into a classical model



The issue

Possible responses

I The data are wrong but again there is no indication for thisI The data are non-phonological

but there is a clearinteraction with ‘real’ phonology





The issue

Possible responses

I The data are wrong but again there is no indication for thisI The data are non-phonological but there is a clear

interaction with ‘real’ phonology





The issue

Possible responses


interaction with ‘real’ phonologyI Phonology is wrong

but this throws away decades of solidresults




The issue

Possible responses



results




The issue

Possible responses



resultsI We have to integrate these facts into a classical model



Voicing in Dutch






Voicing in Dutch

Voicing in Dutch

I A number of authors (Avery 1996, Iverson and Salmons2003, Van Oostendorp 2002, fc) have argued onsynchronic and diachronic phonological grounds that thereis a split in the voicing system of Dutch:

I Stops have a ‘Romance’ system of [± voice]I Fricatives have ‘Germanic’ system of [± spread glottis]

and/or length



Voicing in Dutch

Arguments in favour of a stop/fricative split (1)

I Syllabification: voicing for fricatives (but not stops) ispredictable in intervocalic position ([knœf@l] ‘hug’, [hø:v@l]‘hill’, *[knø:f@l], *[hœv@l]; [kAb@l] ‘ripple’, [ka:b@l] ‘cable’,[kEp@l] ‘yarmulka’, [ke:pi] ‘kepi’)

a. σ σ

@@�� @@��

k n œ f @ l

b. σ σ

�� @@��h ø: v @ l

c. *σ σ

@@��

�� @@��k n ø: f @ l

d. *σ σ

�� @@��h œ v @ l



Voicing in Dutch

Arguments in favour of a stop/fricative split (2)I Voicing assimilation. In obstruent clusters C1C2:

I If C2 is a stop, the cluster gets the underlying voicing of C2(/hœyz/+/dø:r/→[hœzdø:r]) (’house+door’=’‘front door’)

I If C1 is a fricative, the cluster gets devoiced:(/hœyz/+/vœyl/→[hœsfœyl]) (’house+dirt’=’‘garbage’)

I Exceptions. Some dialects have exceptions to finaldevoicing. If these exceptions are lexical, they alwaysinvolve stops only; if they are grammatical they alwaysinvolve fricatives only.

I Spelling. Final devoicing is reflected in traditional Dutchorthography for fricatives (huis-huizen ‘house(s)’) but notstops (bord-borden ‘plate(s)’)



Voicing in Dutch

Arguments in favour of a stop/fricative split (3)

I Phonetics. “The problem is that fricative geminates arealways realized as voiceless, independently of theircontext, exact duration, etc.” (Ernestus 2000:177)



Voicing in Dutch

Voiceless fricatives as long/spread glottis typologically

I VAUX: Fricative ⊃ [spread glottis]. ‘Fricatives preferablyhave the feature [spread glottis]’ (Vaux 1998, Avery 2001)

I MULTILINK: [spread glottis] ⊃ µµ ‘The feature [+spreadglottis] has to be linked to two positions.’ (Ringen 1999)



Voicing in Dutch

Different representations for stops and fricatives/t/ /d/

[-cont]

[coronal]

[-cont]

[coronal] [voice]

/s/ /z/[+cont]

[coronal] [sp.gl.]

[+cont]

[coronal]



Voicing in Dutch

Different representations for stops and fricatives/t/ /d/

[-cont]

[coronal]

[-cont]

[coronal] [voice]

/s/ /z/

[+cont]

[coronal]

µ

[+cont]

[coronal]



Voicing in Dutch

Proposal about Lexicon Optimisation

I Choose the lexical representation which is least marked(contains the smallest amount of structure) and compatiblewith the data

I In case of stops, this will be the voiceless variant; in thecase of fricatives, this will be the variant which is short/not[spr.gl]

I Notice that this makes predictions about e.g. German, inwhich plosives are also characterized by [sp.gl]



Voicing in Dutch

Baayen and Ernestus (2003)

Responses ending in -de and -te, by type of stem-finalobstruent (CV:C words)

TYPE de teP 3 97T 4 96S 55 45F 64 36X 81 13



Voicing in Dutch

The length of place

I This accounts for the split between plosives andobstruents. What about place?

I At least for velarity, we know that there is an intimateconnection to length as well

I Velars sometimes behave as if they were long. E.g.velarisation in codas triggers shortening of precedingvowels.

I [sxu:n@] ∼ [sxuN] ‘shoe(s)’ (Antwerp Dutch)I [zi:t] → [zik] ‘time’ (Cologne German)



Conclusions






Conclusions

Conclusions

I At first sight, Final Devoicing is a very simple andstraightforward phonological process

I Recent empirical study has shown that FD is morecomplicated than was hitherto assumed

I However, this is a reason to refine our models rather thanreject them


voicing variability and formal phonology - stanford...

Documents