Salle Celan, 45 rue d'Ulm
Why do people of different linguistic background sometimes perceive the
same acoustic signal differently? For instance, when hearing nonnative
speech that does not conform to sound structures allowed in their native
language, listeners may report hearing vowels that are not acoustically
present. This phenomenon, known as perceptual vowel epenthesis, has been
attested in various languages such as Japanese, Brazilian Portuguese,
Korean, and English. The quality of the epenthesized vowel varies
between languages, but also within languages, given certain phonemic
How much of this process is guided by information directly accessible in
the acoustic signal? What is the contribution of the native phonology?
How are these two elements combined when computing the native percept?
Two main families of theories have been proposed as explanations:
two-step and one-step theories. The former advocate an initial parsing
of the phonetic categories, followed by repairs by an abstract grammar
(e.g., epenthesis), while one-step proposals posit that all acoustic,
phonetic, and phonological factors are integrated simultaneously in a
probabilistic manner, in order to find the optimal percept.
In this dissertation, we use a combination of experimental and modelling
approaches in order to evaluate whether perceptual vowel epenthesis is a
two-step or one-step process. In particular, we investigate this by
assessing the role of acoustic details in modulations of epenthetic
In a first part, results from two behavioural experiments show that
these modulations are influenced by acoustic cues as well as phonology;
however, the former explain most of the variation in epenthetic vowel
responses. Additionally, we present a one-step exemplar-based model of
perception that is able to reproduce coarticulation effects observed in
human data. These results constitute evidence for one-step models of
nonnative speech perception.
In a second part, we present an implementation of the one-step proposal
in Wilson et al. (2013) using HMM-GMM (Hidden Markov Models with
Gaussian Mixture Models) from the field of automatic speech recognition.
These models present two separate components determining the acoustic
and phonotactic matches between speech and possible transcriptions. We
can thus tweak them independently in order to evaluate the relative
influence of acoustic/phonetic and phonological factors in perceptual
vowel epenthesis. We propose a novel way to simulate with these models
the forced choice paradigm used to probe vowel epenthesis in human
participants, using constrained language models during the speech
In a first set of studies, we use this method to test whether various
ASR systems with n-gram phonotactics as their language model better
approximate human results than an ASR system with a null (i.e., no
phonotactics) language model. Surprisingly, we find that this null model
was the best predictor of human performance.
In a second set of studies, we evaluate whether effects traditionally
attributed to phonology may be predictable solely from acoustic match.
We find that, while promising, our models are only able to partially
reproduce some effects observed in results from human experiments.
Before attributing the source of these effects to phonology, it is
necessary to test ASR systems with more performant acoustic models. We
discuss future avenues for using enhanced models, and highlight the
advantages of using a hybrid approach with behavioural experiments and
computational modelling in order to elucidate the mechanisms underlying
nonnative speech perception.