The lexicon could support unsupervised vowel category learning in Spanish: a new model tested on a new, hand-annotated corpus of Spanish infant-directed speech

Swingley, D. & Alarcon, C.

University of Pennsylvania

The phonetics of vowels are notoriously messy, yet the consensus view
holds that unsupervised distributional clustering would suffice for
infant vowel-category learning. Acoustic measurement of
infant-directed speech suggests otherwise. One solution is to
incorporate contextual information from the lexicon, which infants
begin learning contemporaneously with speech-sound categories
(Bergelson & Swingley, 2012; Feldman et al., 2013; Swingley, 2009).
Prior models suggesting benefits of lexical guidance tested simulated
Gaussian phonetic data with no word- or context-specific phonetics,
thus dodging perhaps the theory's main potential flaw.

We annotated 2700+ Spanish infant-directed vowels, measuring and
hand-checking all boundaries and formants. Because vowel categories
overlap phonetically, unsupervised classification of vowel tokens was
grotesquely ineffective, as predicted, even for Spanish's 5-vowel
system. To test the lexical hypothesis, we implemented unsupervised
clustering over lexical prototypes (means of words' formant values).
Clusters formed over the lexicon were good ("good" means: solutions
where for each vowel (e.g. /e/), the discovered category contained
85+% of that vowel's instances, and little else: 85+% /e/.) The
model's parameter space was validated broadly (minimum word frequency,
number of categories, etc.) and tested under varying assumptions about
word representation. The results provide unique new evidence that
word learning could support infants' phonetic learning.