ultrasound-based tongue root imaging and measurement james m scobbie qmu with thanks to...

38
Ultrasound-based tongue root imaging and measurement James M Scobbie QMU With thanks to collaborators Jane Stuart-Smith, Marianne Pouplier, Alan Wrench, Eleanor Lawson, Olga Gordeeva

Upload: liberty-seel

Post on 14-Dec-2015

214 views

Category:

Documents


0 download

TRANSCRIPT

Ultrasound-based tongue root imaging and measurement

James M Scobbie

QMUWith thanks to collaborators

Jane Stuart-Smith, Marianne Pouplier, Alan Wrench,

Eleanor Lawson, Olga Gordeeva

2Introduction

• Pros and cons of Ultrasound Tongue Imaging

• EPG/UTI experiment on English /l/– Alveolar contact or vocalisation– Light and dark allophones of /l/

• The ECB08 UTI corpus– Scottish derhoticisation and articulation of /r/– Vowel system– A handful of /l/ again…

• Demo of AAA software

3UTI

• From qualitative “transcription” to quantitative laboratory-based studies with stabilisation

4Pros and cons of UTI

• Pro– Tongue root to blade in one image– Instant, real-time, easy, safe, cheap– Qualitative and quantitative analysis– Can be combined with other techniques

• Con– Image quality is variable– Hardly any constriction or info on passive articulator– Frame rate of video output is only ~30Hz (~33ms)– Synchronisation with acoustics is problematic– Quantitative analysis is time-consuming and as yet

poorly developed… what to measure?

5Future: corrected high speed data

6English /l/

• /l/ is lighter in onset, darker in coda

• Many accents have “vocalisation” in coda

• EPG + UTI study of 10 speakers– UTI image quality uniformly awful – EPG results very interesting – Context was /i/+/l/ (+ {/b/, /h/, /l/}) +/i/

• Pee leewards, peel beavers, peel heaps of, etc.

• EPG results– Reduction or loss of alveolar contact in codas– Reduced palatal contact (compared to /i/) due to /l/

7Example onset

• Alveolar contact in orange, palatal in green• S2 typical in losing palatal contact in onset

(can we pee leewardin a gentlebreeze)

8

• No alveolar contact, more palatal contact

(can we peel BBC advertisingfrom the shop window)

Example coda_b retraction

9EPG results: loss of palatal contact

• E + S1: light onset and dark coda in palatality

• Scots S2,3,4 show darker (less [i]-like) onset

• Question 1: what about intergestural timing?

• Question 2: what about the pharyngeal aspect of darkness rather than loss of palatality?

0%

20%

40%

60%

80%

100%

S1 S2 S3 S4Speaker

Co

nta

ct i

-Zo

ne

Onset gemambi Coda_hCoda_b

0%

20%

40%

60%

80%

100%

E1 E2 E3 E4 E5Speaker

Co

nta

ct

i-Z

on

e

Onset gemambi Coda_hCoda_b

10EPG results: timing

• Relatively simultaneous alveolar contact and loss of palatality in onset

• Alveolar contact is delayed in coda (or missing) and loss of palatality occurs earlier

-60

-40

-20

0

20

40

60

80

100

120

onset gem ambi coda_h coda_b

context

La

g (

ms)

S

E

11Coda = vocalised and darker

12UTI: Scottish pharyngealisation?

• Measurement of Tongue root retraction in [i] and in [l] for a single sample speaker S2– Coping with terrible quality UTI

• Find frames of maximum advancement and maximum retraction of root just above hyoid shadow)– Typical problems in measuring images

13Example onset

• Poor image quality

• Time and location of root: top of hyoid shadow

14Example onset

• This is only a bit better than guessing, but impression is of slight pharyngealisation

15Results (S2, n = 18)

• Tongue root retracts earlier in coda_b (p<0.01)– Max advancement appears to be near end of [i] vowel in

onset condition and mid-way through [i] in coda_b condition– Max retraction apparently at end of [l] in onset condition and

towards the end of [b] in coda_b condition

• [i] is less advanced in coda_b than onset (p<0.005)– There is a n.s. trend for greater pharyngealisation in coda [l]

retraction duration

0

50

100

150

200

onset coda_b

ms

retraction distance

3

3.5

4

4.5

5

i lm

m

onset

coda_b

16Conclusions

• Darkness as measured by decrease in palatality in /i/ context shows onset/coda differences for only some speakers– Probably dialectal: Scots /l/ is less [i]-like in onset

• “All” speakers show a strong timing difference– Front and back gestures dissociate in coda so that

posterior gesture is earlier and alveolar (if present at all in coda) is later (“gestural dissociation”)

• Qualitative (and quantitative) analysis of UTI data probably shows greater pharyngealisation for all speakers’ coda than onset.

17ECB08

• Ultrasound/acoustic corpus – 15 teenagers (12-14) in friendship pairs (+4 11yrs)– Wordlist and some spontaneous discourse– Half from a WC and half from a MC school– Main purpose to test effect of use of UTI on

vernacular speech variables

• Secondary purpose– Derhoticisation of coda /r/ - pharyngealisation?– Vowel space

• But sadly not much room for– Vocalisation of coda /l/ - pharyngealisation?

18Derhoticised coda /r/

Hiya my name's Kaj McInally

My company's FinesseDecor (Scotland) Ltd

I'm not a manager. I'm a painter and decorator

to trade, first and foremost

who just so happened to start work for myself, and then

we’ve been that... kinda... successful that we've had to take on people

19Losing /r/ in Scotland

• Since the 1970s coda /r/-“loss” has been reported in working class speech– Not the RP-like middle-class non-rhoticity

• Stuart-Smith (2003) in a Glasgow corpus including 14-15 year old children showed that WC girls have no overtly rhotic consonant for coda /r/ in approximately 90% of cases, boys in about 80%– Middle class children and older adults are rhotic, so the

stratified derhoticisation is indicative of change in progress.– /r/ seems to be turning into a vowel right now– Strong impression of pharyngealisation offglide on vowels

with monophthongal pharyngealisation on low back ones

20

rain, with an anterior approximant, usually described as being “retroflex” (note low F3)

ferry, with a tap (an approximant is more common)

F3F2

Typically rhotic tokens of Scottish /r/

21Word-final derhoticisation in ECB08

Rhotic ear (above) car (below)

F3F2

F3F2

F3F2

F3F2

Derhoticised ear (above) car (below)

22Rhotic (MC) speakers

• Lexical sets BIRD WORD HERD merged (8/11)– Earth, verb, berth, (err) = third, word, surf, birth, fur– Could be a rhotic vowel /ɚ/

• No /a/ split (Pam/palm are homophones)

• /ʉ/ is central and not very high

iɹ ʉɹ oɹ

eɹ ɚ ɔɹ

ɑɹ

i ʉ o

e ı ɔ

ɛ a ʌ

23MC Edinburgh

ɔ

ɛ

ɪ

ʌ

a

e

io

u

1300

300

Dim

ensi

on 1

3300 800Dimension 2

F1

(Hz)

F2 (Hz)

24WC West Lothian

ɔ

ɛ ɪ

ʌ

a

e

i o

u

1300

300

Dim

ensi

on 1

3300 800Dimension 2

F1

(Hz)

F2 (Hz)

25Articulation of vowels (EF4)

• Phonologically, only /ɛ I / are “lax”

26Sample ultrasound images of /r/

• Tipup (LM17 onset) or tipdown (LM15 onset)

27Waterfall time sequence: hair

Tongue root retraction

Tongue blade raising

[he]

[he]

[ɹ]

[ɹ]

[ə]

28

iə ʉə oʌ

eə ɔˤ

ɛˤ ɑː (ʕ)

• More vowels (and environments) with weak /r/– No merger of /ɛr/ and /ʌr/ (8/8)– /a/ split (hat/heart) [a] vs. [ɑ] for the most derhotic– /ʌr/ is short without compensatory lengthening– High vowels create diphthongs– Pre-pausal /r/ tends to devoice

• Potential /ʌ/ merger (hut/hurt, bud/bird)

i ʉ o

e ı ɔ

ɛ a ʌ

Derhoticising (WC) speakers

29

Pre-pausal /r/ may have late (covert?) tip

car

• Low vowels sound derhoticised, acoustically lack F2/F3 approximation, and are near-monophthongs.

• Articulatorily a clear rhotic gesture was retained

30

Covert rhoticity occurs even in weak syllables and in spontaneous speech

31

• What about /l/?– If dark, is it pharyngealised?– If vocalised, is it a pharyngeal?– How are derhoticised /r/ & vocalised /l/ kept apart?– Hip hum hut– Fur/fir hurt– Pill film– Mull bulb cult

• Clear difference between /r/ and /l/ in open and closed syllables

/l/ in a derhoticising (WC) speaker

32UTI of laterals

– Red = // mull (cons) & bulb (vocalised)– Blue = /ı/ film (cons) & pill (vocalised)

• Pharyngealisation vs. velarisation?

33UTI of laterals

– Red = cult (cons /lt/)– Green = hurt (cons /t/)

• /l/ pharyngealised + velarised?

• Pharyngealised postalveolar /r/ with saddle

34/l/ compared to /o/ and /ɔ/

• Pharyngealisation and velarisation more extreme than in vowels

35Conclusions

• Onset/coda differences in /l/ in a high vowel context are well-known to involve loss of palatality and a greater pharyngeal constriction (Sproat and Fujimura 1993), plus subtle loss of alveolar contact (eg Giles & Moll)

• Scottish speakers who have no onset/coda difference in palatality do show increased pharyngealisation in coda (and may show very strong vocalisation, not gestural undershoot)

• Vocalised /l/ may be velarised while pharyngealisation occurs for consonantal /l/

36Conclusions

• Derhoticisation often sounds like pharyngealisation

• But in prepausal and other masking contexts there can be delayed covert post-alveolar constriction, due to “gestural dissociation”

• WC /r/ seems to be changing from consonant into vowel, with some increase in vowel space

• Meanwhile, MC rhotic speakers merge vowels• WC /l/ and /r/ seem to be keeping distinct

– Is the pharyngealised /l/ also velarised?– Is the difference purely anterior?

37AAA demo

• Let’s look at pharyngealisation in a derhoticising speaker – Hut vs. hurt– Bud vs. bird– Far vs. fir

38Who says you need ultrasound?