A longitudinal study of the oral properties
of the French-English interlanguage

A quantitative approach of the acquisition
of the /ɪ/-/iː/ and /ʊ/-/uː/ contrasts

Adrien Méli

April 4, 2018

Why and how my work evolved
Is the ℋ₀ accepted?
What remains to be done…

Why and how my work evolved

The null hypothesis (ℋ₀)

There are no differences in the phonological acquisition of:

The initial data (1/2)

More than 14 hours of recorded conversations over 4 sessions

(9 three-session and 9 four-session female speakers + 3 four-session male speakers)

The initial data (2/2)

Text in French recorded by the four-session speakers
List of English words recorded by the four-session speakers
“Homemade” corpus of native spontaneous speech (NSS, ~4.5K tokens)

The final, generated data (1/2)

A multitier Praat TextGrid with phonemic and syllabic intervals aligned by SPPAS and P2FA for each recording

DID0024

DID0168

The final, generated data (2/2)

Datasheets with information for all corpora (1 row = 1 vowel)

tibble

WHY all this?

Each field has specificities…

… which taken together, are at odds with one another.

(Add to this, the longitudinal aspect of the study)

So a new approach had to be devised

Columns Columns

Theory-driven: BEFMOA, BEFPOA, CLXFREQ…
Data-driven: PHONBEF, PHONAFT, PHONAFT…
Opportunistic and skill-driven: xSYLL, EPENTHETIC, Fxxx…

Is the ℋ₀ accepted?

The counts of phonemes are unevenly distributed

(n = 17,189 for the 12 four-session speakers)

Does the analysis depend on how the data is processed?

A look at the mean Euclidean distances
between female native and learner values in the F1/F2 vocalic space

For mid-temporal values, the differences
between raw and BDM-normalized values are minimal

Inter-speaker comparisons:
A trend for /ɪ/?

When taking Peterson & Barney as reference point
the distances are greater than with the NSS values

When taking Peterson & Barney as reference point
/ɪ/ and /iː/ seem closer to native values

Reminder: Discrete Cosine Transform

(x-axis: percentage of the vowel’s duration; y-axis: Bark)

DCT principle

Discrete Cosine Transforms: K0
Inter-phoneme differences

Discrete Cosine Transforms: K0
Inter-speaker differences

Discrete Cosine Transforms: K1
Inter-phoneme differences

Discrete Cosine Transforms: K1
Inter-speaker differences

A tentative conclusion

Regardless of how the data is processed,
/ʊ/ and /uː/ feature more instability,
and their formant values are further
from native values than those of /ɪ/ and /iː/.

The ℋ₀ can be rejected.

What remains to be done

Treating lexical differences

Lexical distributions

Refining classifications

Example of a KNN-based confusion matrix: