📄 Coordinations spatio-temporelles dans les suites ab(b)i en arabe marocain (Spatio-temporal coordinations in Moroccan Arabic ab(bi) sequences) [in French]

Summary

The paper asks how Moroccan Arabic geminate /bb/ differs from singleton /b/ in timing and space, and how the consonant is timed against the vowels around it. Two adult Moroccan speakers were recorded with 3D electromagnetic articulography (AG500 Carstens Medizinelektronik). Each produced the minimal pair [nsabi] ('my sons-in-law') and [tsabbi] (a 2nd-person feminine imperative, 'scold') eight times inside a fixed carrier sentence. Lower-lip and tongue sensor traces were segmented automatically from velocity profiles, and acoustic durations were also measured.

Compared with /b/, the geminate showed a longer lip gesture overall, a much longer closure plateau and a higher lip target. The lip closure also started earlier in [abbi], which appears to explain why the preceding [a] is shorter there. The articulatory interval between the [a] and [i] targets was longer in [abbi].

This pattern argues against Öhman's model, in which consonants are simply laid over an independent vowel-to-vowel movement. It fits Articulatory Phonology, where consonant and vowel gestures are timed relative to each other. It also fits the view of a geminate as two coordinated identical gestures.

Key points

• Data: 2 speakers (S1, S2), 8 repetitions, target items [nsabi] vs. [tsabbi] in the frame [ibi___hnaa] ('bring ___ here'), with stress on [a]. Recordings were sampled at 200 Hz with sensors TTIP, TMID, TDOR, LLIP and JAW, and landmarks were extracted with Mview (M. Tiede, Haskins).
• Landmark thresholds: consonant onset, target, release and offset were set where velocity reaches 20% of peak. Vowel landmarks used a 10% threshold, on TDORx for S1 and TMIDy for S2.
• Acoustic durations: /bb/ lasted 126 vs. 73 ms for S1 (+73%) and 119 vs. 64 ms for S2 (+86%). [a] was 13–14% shorter before /bb/, which is below Rossi's 20% differential threshold, so it is probably not a perceptual cue.
• Lip gesture: total duration rose by about 25–26% (164→205 ms and 155→196 ms). The plateau rose by 73.6% and 62%. The lower lip sat significantly higher during /bb/, in line with Löfqvist's (2005) "higher virtual target" idea.
• Timing: the [a]–[i] interval was longer in [abbi]. The time from the [a] target to the consonant midpoint stayed stable across the two words, and the gesture ended at about the same point relative to [i].

Interesting ideas

• I1. All measurements are aligned to one shared anchor event (the [a] target) before comparing conditions, so timing differences can be read on a common time axis.
• I2. The study is built around a single quantity on which two rival models make opposite predictions (constant vs. longer [a]–[i] interval), so one measurement can discriminate between them.
• I3. A statistically significant difference (the 13–14% vowel shortening) is checked against a perceptual threshold before being treated as meaningful.
• I4. Landmarks are defined relative to each movement's own peak velocity, not by absolute values, which makes automatic segmentation reproducible across speakers.
• I5. The stimuli are a tightly controlled minimal pair inside a fixed carrier sentence, and statistics are run per speaker to avoid pooling variability.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (world): event-ordering could place story events on a timeline relative to one anchor event, making "before/after/overlap" claims comparable and explainable across texts.
• I2 (learn): when two witness parsers or repair rules disagree, choose evaluation items where their predictions diverge, rather than averaging over sentences where both agree.
• I3 (learn): evaluation reports could require gains to exceed a pre-set practical margin, not just pass a significance test, before a model change is accepted.
• I4 (learn): confidence cut-offs in silver-data agreement could be set relative to each parser's own score distribution instead of one global absolute threshold.
• I5 (tree): probe the parser with minimal-pair sentences in a fixed frame and score per treebank or genre separately, to isolate exactly which change shifts the attachment.