2. Conditioned allophony in speech perception: An MEG study Mary Ann Walter & Valentine Hacquard walterma@mit.eduhacquard@mit.edu Dept. of Linguistics & Philosophy, MIT KIT/MIT MEG Lab 2 nd Old World Conference in Phonology CASTL-Tromsø, January 20-22, 2005  

3. Questions about (conditioned) allophony: Structural properties of a language’s phonology significantly affect the perception of speech sounds (particularly well-documented wrt phonemic inventory (Kuhl 1993, Best 1995) ) How perceptible are differences between allophones to speakers? How does the perceived similarity of allophones compare to that of phonemes, free variants, and other kinds of contrast? What implications follow for models of phonology and similarity computation?

4. Overview: Subphonemic perceptibility MEG and the MMF Where allophones fit in: Russian/Korean French/Spanish Quebecois French Results: Equivocal, but ours suggest that allophones affiliate with structural, phonemic contrasts, rather than within-category free variants

5. Allophone perceptibility: Speakers produce consistent and finely controlled distinctions between allophones Suggests they must be able to perceptually distinguish them, at least proprioceptively Anecdotally this often seems not to be the case When measured experimentally, subjects typically distinguish allophones at above chance levels, but much less easily and well than phoneme pairs (for aspiration in English, see Pegg & Werker 1997, Whalen et al. 1997, Utman et al. 2000, Jones 2001) For the segment pair we will discuss (e and ε), evidence is contradictory: Pallier et al (1997) find that the pair is indistinguishable for Spanish speakers for whom they are conditioned allophones, even when bilingual in a language in which they are phonemic (Catalan) Escudero and Boersma (2002) find that discrimination does improve for such speakers, however

6. Magnetoencephalography (MEG): Measures brain responses as indexed by magnetic activity Millisecond-by-millisecond temporal resolution Gradient responses Avoids problems and task effects of offline similarity judgment tasks Consistency of rating scale Understanding of task 156 channels

7. Origin of signal in auditory cortex The mismatch field response (MMF): An non-attentive auditory response that indexes perceived (cf tone experiments) (dis)similarity One of a number of language-sensitive neural responses Peaks between 180-250 ms post stimulus onset

8. The mismatch field response (MMF): Each line represents averaged sensors The mismatch response is the difference between the standard and deviant responses at peak/in the relevant time window

9. The mismatch field response (MMF): Voicing (Sharma & Dorman 1999) VOT (ms) 010203040506070 dt dt t1t1 t2t2 Significant MMF for both pairs (  within-category differences perceptible) Significantly greater MMF for the cross-category pair, despite equal acoustic distance

10. The mismatch field response (MMF): Vowel backness (Näätänen et al. 1997) F2 (Hz) In general, larger frequency deviation  larger MMF For Finns, significantly greater MMF for deviant ö than õ, despite greater acoustic distance of latter from standard (e) e 1940 o 851 õ 1311 ö 1533 Finnish + Estonian Finnish + Estonian Finnish + Estonian

11. Studies have examined contrasts of: phonemic pairs versus within-category free variants (Sharma & Dorman) phonemic pairs versus non-prototypical segments (Näätänen et al.)  Phonemic contrasts have a special status in speech processing Are they the only ones? Allophonic contrasts are also encoded in a language’s phonology Allophones show a bimodal distribution on the surface, like phonemes Contrast types and the MMF:

12. Predictions: Phonemes a-e a-e a-e Allophones e-ε e-ε e-ε Free variantsε 1 - ε 2 ε 1 - ε 2 ε 1 - ε 2 MMF Amplitude

13. Allophonic mismatches: Russian/Korean Voicing (Kazanina & Phillips 2004) Russian: d/t phonemically distinct Phonemes elicit a significant MMF response; allophones not at all Korean: d/t distributed allophonically (intervocalic voicing) ?

14. Predictions: Phonemes a-e a-e a-e Allophones e-ε e-ε e-ε Free variantsε 1 - ε 2 ε 1 - ε 2 ε 1 - ε 2 Kazanina & Phillips 2004 MMF Amplitude

15. e vs.  French: phonemic contrast Spanish (Buenos Aires): allophonic contrast e  / __ C ] σ Spanish (Puerto Rican): free variation Allophonic mismatches: French/Spanish vowel tenseness – a three-way comparison

16. e e e e e e e a e e e e  e e e e e e e  e e e Standard:/e/1050 Deviants:/a/100random order /  / 100 Non-attentive paradigm (silent movie) French (n=10), Spanish-Argentinian (n=9), Spanish-Puerto Rican (n=4) Method:

17. Results: 0 1E-14 2E-14 3E-14 4E-14 5E-14 6E-14 FAPR T E e a Middle bars represent baseline response to standard /e/, flanking bars are responses to deviants /  / (left) and /a/ (right) Mismatch response is significant for both deviants for all language groups (p<.02) (contrast with Kazanina and Phillips’ findings)

18. Results: A downward progression is observed in the MMF response, according to language group However, these differences do not reach statistical significance A trend differentiating French and Argentinian Spanish from Puerto Rican Spanish appears (p=.11)

19. Predictions: Phonemes a-e a-e a-e Allophones e-ε e-ε e-ε Free variantsε 1 - ε 2 ε 1 - ε 2 ε 1 - ε 2 Kazanina & Phillips 2004 MMF Amplitude ?

20. * * T x 10^-13 Results: a-e contrast elicits greater MMF for all language groups greater acoustic distance consistently phonemic Significant for French and Puerto Rican, not Argentinian (p<.001, p=.043 vs p=.095)

21. Results: For Argentinian speakers the a-e and e-ε contrasts are both phonologically relevant, and therefore qualitatively of the same type  MMF amplitude not significantly different For Puerto Rican speakers one is phonologically relevant (a-e) and one is not (free variants e-ε) – the contrasts are of qualitatively different types  MMF amplitudes are significantly different For French the responses also differ, though both contrasts are phonemic French MMF responses greater across the board (inventory size effect) (Hacquard & Walter 2003) Differences also spread wider and reach significance sooner

22. Results: Why the inconsistencies between these two studies? K&P manipulate VOT in stimuli  not a primary cue for voicing in Korean K&P include multiple tokens in each category  biases subjects toward focusing on categorically phonemic distinctions Use of consonants versus vowels  different timing of acoustic information results in different processing

23. Experiment 2 (in progress): Quebecois French – a within-language three-way comparison High vowels alternate between long/tense and short/lax (Dumas 1976, Dechaine 1991, Martin 2002) word-finalword-initialword-medial open syllablelonglong/shortshort/devoiced/deleted closed syllableshortshortshort/devoiced iuIUεiuIUε free variants phonemic allophonic Stay tuned!

24. Conclusions: Allophones are distinguishable in early speech processing Allophones appear to pattern with other structural, phonological contrasts, in contrast to within-category free variants In conjunction with behavioral research on the relative (im)perceptibility of allophone pairs, these results necessitate a model of phonology in which similarity may be computed over at least three domains: phonemic, allophonic, and acoustic

25. Thank you – now let’s look at the sun!