Effects of phonological contrast on phonetic variation in Hindi and - - PowerPoint PPT Presentation

Effects of phonological contrast on phonetic variation in Hindi and English stops

Ivy Hauser University of Massachusetts Amherst

blogs.umass.edu/ihauser

Workshop on Phonological Variation and its Interfaces November 22, 2018

Introduction

Big picture: What is the typology of variation? How does variation itself vary across languages? What are the limits?

1

Introduction

Big picture: What is the typology of variation? How does variation itself vary across languages? What are the limits?

◮ What varies 1

Introduction

Big picture: What is the typology of variation? How does variation itself vary across languages? What are the limits?

◮ What varies ◮ Amount of variation 1

Introduction

Big picture: What is the typology of variation? How does variation itself vary across languages? What are the limits?

◮ What varies ◮ Amount of variation ◮ Sources of structure in variation 1

Introduction

→ Question: How do different systems of phonological

contrast affect patterns of phonetic variation?

2

Introduction

→ Question: How do different systems of phonological

contrast affect patterns of phonetic variation?

◮ Hypothesis: “Systems with more phonological contrasts

should exhibit less within-category variation than systems with fewer contrasts” (Lindblom, 1986).

2

Introduction

→ Question: How do different systems of phonological

contrast affect patterns of phonetic variation?

◮ Hypothesis: “Systems with more phonological contrasts

should exhibit less within-category variation than systems with fewer contrasts” (Lindblom, 1986).

◮ Hypothesis: Variation predicted by number of phonemes in

an inventory.

2

Introduction

→ Question: How do different systems of phonological

contrast affect patterns of phonetic variation?

◮ Hypothesis: “Systems with more phonological contrasts

should exhibit less within-category variation than systems with fewer contrasts” (Lindblom, 1986).

◮ Hypothesis: Variation predicted by number of phonemes in

an inventory.

◮ But phonological contrasts are not unidimensional in

phonetic space

2

Introduction

→ Question: How do different systems of phonological

contrast affect patterns of phonetic variation?

◮ Hypothesis: “Systems with more phonological contrasts

should exhibit less within-category variation than systems with fewer contrasts” (Lindblom, 1986).

◮ Hypothesis: Variation predicted by number of phonemes in

an inventory.

◮ But phonological contrasts are not unidimensional in

phonetic space

◮ Issues with quantifying within-category variation:

What are the relevant phonetic dimensions? What counts as a system/inventory?

2

Introduction

Proposal: We expect to see less variation in languages that realize more phonological contrasts only along the particular phonetic dimensions which realize additional contrasts.

3

Test case: Hindi and English stops

Hindi has four contrasting stops at each place of articulation; English has two (Kagaya and Hirose, 1975; Ohala, 1994; Quirk et al., 1972). Labial Coronal Retroflex Velar Hindi /p /ph/ /b/ /bh/ /t/ /th/ /d/ /dh/ /ú/ /úh/ /ã/ /ãh/ /k/ /kh/ /g/ /gh/ English /p/ /b/ /t/ /d/ /k/ /g/

4

Test case: Hindi and English stops

◮ If variation is predicted by number of phonemes in an

inventory:

5

Test case: Hindi and English stops

◮ If variation is predicted by number of phonemes in an

inventory: We might expect Hindi speakers to constrain variation on all dimensions, including lag time, prevoicing, f0, etc.

5

Test case: Hindi and English stops

◮ If variation is predicted by number of phonemes in an

inventory: We might expect Hindi speakers to constrain variation on all dimensions, including lag time, prevoicing, f0, etc.

→ Hindi /kh/ should vary less than English /kh/ in voiceless lag time.

5

Test case: Hindi and English stops

◮ If variation is predicted by number of phonemes in an

inventory: We might expect Hindi speakers to constrain variation on all dimensions, including lag time, prevoicing, f0, etc.

→ Hindi /kh/ should vary less than English /kh/ in voiceless lag time.

◮ If variation is predicted by more contrasts along a single

dimension:

5

Test case: Hindi and English stops

◮ If variation is predicted by number of phonemes in an

inventory: We might expect Hindi speakers to constrain variation on all dimensions, including lag time, prevoicing, f0, etc.

→ Hindi /kh/ should vary less than English /kh/ in voiceless lag time.

◮ If variation is predicted by more contrasts along a single

dimension: Hindi speakers will only exhibit less variation along phonetic dimensions which distinguish additional contrasts relative to English.

5

Test case: Hindi and English stops

◮ If variation is predicted by number of phonemes in an

inventory: We might expect Hindi speakers to constrain variation on all dimensions, including lag time, prevoicing, f0, etc.

→ Hindi /kh/ should vary less than English /kh/ in voiceless lag time.

◮ If variation is predicted by more contrasts along a single

dimension: Hindi speakers will only exhibit less variation along phonetic dimensions which distinguish additional contrasts relative to English.

→ Hindi /k/ and /g/ should vary less than English /g/ in voicing.

5

Phonetic dimensions in Hindi and English stops

Hindi lag time closure voicing /k/ /kh/ /g/ /gh/ English lag time closure voicing /g/ /k/

6

Phonetic dimensions in Hindi and English stops

Hindi lag time closure voicing /k/ /kh/ /g/ /gh/ English lag time closure voicing /g/ /k/

◮ No difference expected in voiceless lag time (positive

VOT).

◮ Difference expected in prevoicing because Hindi has

additional voicing contrasts relative to English.

6

The experiment

◮ Stimuli: C1VC2 words in carrier phrases. ◮ C1 was a stop and V was one of [i a u]. ◮ Carrier phrases: Say X again; Dobara X doharao. 7

The experiment

◮ Stimuli: C1VC2 words in carrier phrases. ◮ C1 was a stop and V was one of [i a u]. ◮ Carrier phrases: Say X again; Dobara X doharao. ◮ 4 repetitions of each distinct stimulus per speaker. 7

Analysis: Lag time

All data were forced aligned with the Montreal Forced Aligner (McAuliffe et al., 2017) and analyzed in Praat (Boersma and Weenink, 2001)

8

Analysis: Lag time

All data were forced aligned with the Montreal Forced Aligner (McAuliffe et al., 2017) and analyzed in Praat (Boersma and Weenink, 2001) Voiceless stops: voiceless lag time measured from the burst to the

• nset of voicing with AutoVOT (Keshet et al., 2014) and manual

correction.

8

Analysis: Lag time

All data were forced aligned with the Montreal Forced Aligner (McAuliffe et al., 2017) and analyzed in Praat (Boersma and Weenink, 2001) Voiceless stops: voiceless lag time measured from the burst to the

• nset of voicing with AutoVOT (Keshet et al., 2014) and manual

correction.

8

Expected results: Lag time

0.00 0.01 0.02 0.03 0.04 −50 −25 25 50

centered lag time (ms) density Language

English Hindi

Expected results: inventory prediction

density

All graphing done in R (R Core Team, 2013; Wickham, 2009).

9

Results: Lag time

0.00 0.01 0.02 0.03 −50 −25 25 50

centered VOT density Language

English Hindi

Voiceless aspirated coronal VOT values

0.00 0.01 0.02 0.03 −50 −25 25 50

centered VOT density Language

English Hindi

Voiceless aspirated velar VOT values 10

Discussion: Lag time

Why no difference in variation between languages?

(Levene’s Test for homogeneity of variance not significant.)

11

Discussion: Lag time

Why no difference in variation between languages?

(Levene’s Test for homogeneity of variance not significant.)

◮ Voiceless lag time realizes one contrast in both languages,

no difference expected.

11

Discussion: Lag time

Why no difference in variation between languages?

(Levene’s Test for homogeneity of variance not significant.)

◮ Voiceless lag time realizes one contrast in both languages,

no difference expected.

◮ Additional evidence for understanding prevoicing and lag

as separate dimensions (Mikuteit & Reetz, 2007)

11

Discussion: Lag time

voice onset time /g/ /gh/ /k/ /kh/

12

Discussion: Lag time

voice onset time /g/ /gh/ /k/ /kh/ Hindi lag time closure voicing /k/ /kh/ /g/ /gh/

12

Analysis: Voicing

Closure duration was measured from the end of the vowel to the stop burst.

13

Analysis

English voiced stop - prevoiced postvocalically

14

Analysis

English voiced stop - voiceless postvocalically

15

Results: Prevoicing

16

Results: Prevoicing

Prevoicing categories (Beckman et al., 2013) none = less than 25% voicing during closure part = 25-90% percent voiced full = 90%+ voiced

17

Results: Prevoicing

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

All Hindi voiced stops

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

All English voiced stops

18

Examining the English variation

There is more variation in English prevoicing, in accordance with the revised hypothesis.

19

Examining the English variation

There is more variation in English prevoicing, in accordance with the revised hypothesis.

◮ The phonological system of English allows more prevoicing

variation without threatening the maintenance of contrast.

◮ This additional variation is not entirely random – it is

structured by a number of factors.

◮ Between and within-speaker sources of variance.

19

Examining the English variation

There is more variation in English prevoicing, in accordance with the revised hypothesis.

◮ The phonological system of English allows more prevoicing

variation without threatening the maintenance of contrast.

◮ This additional variation is not entirely random – it is

structured by a number of factors.

◮ Between and within-speaker sources of variance. ◮ This structure emerges in English, but not in Hindi.

19

Examining the English variation

There is more variation in English prevoicing, in accordance with the revised hypothesis.

◮ The phonological system of English allows more prevoicing

variation without threatening the maintenance of contrast.

◮ This additional variation is not entirely random – it is

structured by a number of factors.

◮ Between and within-speaker sources of variance. ◮ This structure emerges in English, but not in Hindi.

Sources of structure:

◮ Individual differences ◮ Vowel contexts 19

Individual differences in voicing

20

Individual differences in voicing

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

All Hindi voiced stops

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

Speaker with least voicing

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

Speaker with most voicing

21

Individual differences in voicing

22

Individual differences in voicing

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

All English voiced stops

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

Speaker with least voicing

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

Speaker with most voicing

23

Individual differences: Discussion

◮ Individual speakers exhibit differences in patterns of

prevoicing in English, but not in Hindi.

24

Individual differences: Discussion

◮ Individual speakers exhibit differences in patterns of

prevoicing in English, but not in Hindi.

◮ English differences appear to be unrelated to social factors

(cf. Jacewicz et al., 2009; Herd et al., 2016; Hunnicutt and Morris, 2016)

24

Voicing across vowel contexts

◮ More prevoicing before high vowels has been reported in

English (Smith and Westbury, 1975).

◮ This result replicated here in English but not Hindi.

25

Voicing across vowel contexts

Hindi phonologically voiced stops

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

Hindi stops before [a]

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

Hindi stops before [i]

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

Hindi stops before [u]

26

Voicing across vowel contexts

Hindi phonologically voiced stops

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

Hindi stops before [a]

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

Hindi stops before [i]

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

Hindi stops before [u]

English phonologically voiced stops

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

English stops before [a]

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

English stops before [i]

25 50 75 100 none part full

degree of prevoicing percent of all voiced stops

English stops before [u]

26

Voicing across vowel contexts

◮ Smith & Westbury (1975) proposed an articulatory

explanation:

◮ Easier to continue voicing when vocal folds tighter due to

vowel height.

27

Voicing across vowel contexts

◮ Smith & Westbury (1975) proposed an articulatory

explanation:

◮ Easier to continue voicing when vocal folds tighter due to

vowel height.

◮ Hindi speakers constrain variation despite these

articulatory pressures.

27

Voicing across vowel contexts

◮ Smith & Westbury (1975) proposed an articulatory

explanation:

◮ Easier to continue voicing when vocal folds tighter due to

vowel height.

◮ Hindi speakers constrain variation despite these

articulatory pressures.

◮ Or – the explanation is not articulatory.

27

Sources of variance

◮ We can model amount of voicing in the closure as a

function of several factors.

28

Sources of variance

◮ We can model amount of voicing in the closure as a

function of several factors.

◮ Beta regression for proportion data bounded by (0,1). ◮ Different models for English and Hindi

28

Sources of variance

◮ We can model amount of voicing in the closure as a

function of several factors.

◮ Beta regression for proportion data bounded by (0,1). ◮ Different models for English and Hindi

◮ Differences between languages.

◮ Underlying phonological category accounts for 79% of all

voicing variation in Hindi.

28

Sources of variance

◮ We can model amount of voicing in the closure as a

function of several factors.

◮ Beta regression for proportion data bounded by (0,1). ◮ Different models for English and Hindi

◮ Differences between languages.

◮ Underlying phonological category accounts for 79% of all

voicing variation in Hindi.

◮ ...and 15% of all voicing variation in English.

28

Sources of variance

29

Sources of variance

30

Discussion

◮ Main result: Patterns of phonetic variation are

language-specific.

31

Discussion

◮ Main result: Patterns of phonetic variation are

language-specific.

◮ Relative differences can be predicted by how phonological

contrasts are implemented in phonetic space.

31

Discussion

◮ Main result: Patterns of phonetic variation are

language-specific.

◮ Relative differences can be predicted by how phonological

contrasts are implemented in phonetic space.

◮ Non-contrastive structure emerges when variation does not

threaten contrast maintenance.

31

Discussion

◮ Main result: Patterns of phonetic variation are

language-specific.

◮ Relative differences can be predicted by how phonological

contrasts are implemented in phonetic space.

◮ Non-contrastive structure emerges when variation does not

threaten contrast maintenance.

◮ Regarding Lindblom (1986): The mathematically intuitive

“larger inventory = less variation” hypothesis is not trivially true.

◮ Must be implemented over particular phonetic dimensions.

31

Discussion

◮ Main result: Patterns of phonetic variation are

language-specific.

◮ Relative differences can be predicted by how phonological

contrasts are implemented in phonetic space.

◮ Non-contrastive structure emerges when variation does not

threaten contrast maintenance.

◮ Regarding Lindblom (1986): The mathematically intuitive

“larger inventory = less variation” hypothesis is not trivially true.

◮ Must be implemented over particular phonetic dimensions.

31

Discussion

◮ Main result: Patterns of phonetic variation are

language-specific.

◮ Relative differences can be predicted by how phonological

contrasts are implemented in phonetic space.

◮ Non-contrastive structure emerges when variation does not

threaten contrast maintenance.

◮ Regarding Lindblom (1986): The mathematically intuitive

“larger inventory = less variation” hypothesis is not trivially true.

◮ Must be implemented over particular phonetic dimensions.

◮ Ongoing/future work: more test cases, comparing

non-contrastive to contrastive phonetic dimensions in the same language.

31

Acknowledgements

Thanks to research assistants Greg Feliu and Saumya Joshi. Thanks to Kristine Yu, John Kingston, Joe Pater, Gaja Jarosz, and Sang-Im Lee Kim for input and feedback.

This material is based upon work supported by National Science Foundation Grant no. 1823869 and the National Science Foundation Graduate Research Fellowship Program Grant no. 1451512. Any

• pinions, findings, and conclusions or recommendations expressed in

this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

32

References

Boersma, P ., Weenink, D., 2001. Praat, a system for doing phonetics by computer . Herd, W., Torrence, D., Carino, J., 2016. Prevoicing differences in Southern English: Gender and ethnicity effects. The Journal of the Acoustical Society of America 139, 2217–2217. Hunnicutt, L., Morris, P . A., 2016. Prevoicing and Aspiration in Southern American English. University of Pennsylvania Working Papers in Linguistics 22, 24. Jacewicz, E., Fox, R. A., Lyle, S., 2009. Variation in stop consonant voicing in two regional varieties of American

• English. Journal of the International Phonetic Association 39, 313–334.

Kagaya, R., Hirose, H., 1975. Fiberoptic electromyographic and acoustic analyses of hindi stop consonants. Annual Bulletin, Research Institute of Logopedics and Phoniatrics 9, 27–46. Keshet, J., Sonderegger, M., Knowles, T., 2014. Autovot: A tool for automatic measurement of voice onset time using discriminative structured prediction [computer program]. version 0.91. McAuliffe, M., Socolof, M., Mihuc, S., Wagner, M., Sonderegger, M., 2017. Montreal forced aligner [computer program] version 0.9.0, retrieved from http://montrealcorpustools.github.io/montreal-forced-aligner/. Ohala, M., 1994. Hindi. Journal of the International Phonetic Association 24, 35–38. Quirk, R., Greenbaum, S., Leech, G. N., Svartvik, J., et al., 1972. A grammar of contemporary english . Smith, B. L., Westbury, J. R., 1975. Temporal control of voicing during occlusion in plosives. The Journal of the Acoustical Society of America 57, S71–S71.

33

Appendix: Model output

34

Appendix: Model output

35

Appendix: Model output

36