Word learning relies on the ability to master the sound con-trasts that are phonemic (i.e., signal meaning difference) ina given language. Though the timeline of phoneme develop-ment has been studied extensively over the past few decades,the mechanism of this development is poorly understood. Pre-vious work has shown that human learners rely on referentialinformation to differentiate similar sounds, but largely ignoredthe problem of taxonomic ambiguity at the semantic level (twodifferent objects may be described by one or two words de-pending on how abstract the meaning intended by the speakeris). In this study, we varied the taxonomic distance of pairs ofobjects and tested how adult learners judged the phonemic sta-tus of the sound contrast associated with each of these pairs.We found that judgments were sensitive to gradients in thetaxonomic structure, suggesting that learners use probabilisticinformation at the semantic level to optimize the accuracy oftheir judgements at the phonological level. The findings pro-vide evidence for an interaction between phonological learningand meaning generalization, raising important questions abouthow these two important processes of language acquisition arerelated.

A growing body of research on early word learning suggeststhat learners gather word-object co-occurrence statistics acrosslearning situations. Here we test a new mechanism wherebylearners are also sensitive to word-word co-occurrence statis-tics. Indeed, we find that participants can infer the likely ref-erent of a novel word based on its co-occurrence with otherwords, in a way that mimics a machine learning algorithmdubbed ‘zero-shot learning’. We suggest that the interactionbetween referential and distributional regularities can bring ro-bustness to the process of word acquisition.

The language discrimination process in infants has been suc-cessfully modeled using i-vector based systems, with re-sults replicating several experimental findings. Still, recentwork found intriguing results regarding the difference betweenmonolingual and mixed-language exposure on language dis-crimination tasks. We use two carefully designed datasets,with an additional “bilingual” condition on the i-vector modelof language discrimination. Our results do not show any dif-ference in the ability of discriminating languages between thethree backgrounds, although we do replicate past observationsthat distant languages (English-Finnish) are easier to discrimi-nate than close languages (English-German). We do, however,find a strong effect of background when testing for the abilityof the learner to automatically sort sentences in language clus-ters: bilingual background being generally harder than mixedbackground (one speaker one language). Other analyses revealthat clustering is dominated by speakers information ratherthan by languages.

According to the Language Familiarity Effect (LFE), people are better at discriminating between speakers of their native language. Although this cognitive effect was largely studied in the literature, experiments have only been conducted on a limited number of language pairs and their results only show the presence of the effect without yielding a gradual measure that may vary across language pairs. In this work, we show that the computational model of LFE introduced by Thorburn, Feldman, and Schatz (2019) can address these two limitations. In a first experiment, we attest to this model's capacity to obtain a gradual measure of the LFE by replicating behavioural findings on native and accented speech. In a second experiment, we evaluate LFE on a large number of language pairs, including many which have never been tested on humans. We show that the effect is replicated across a wide array of languages, providing further evidence of its universality. Building on the gradual measure of LFE, we also show that languages belonging to the same family yield smaller scores, supporting the idea of an effect of language distance on LFE.

Adult speech perception is tuned to efficiently process native phonetic categories, causing difficulties with certainnon-native categories. For example, Japanese has no equivalent of the distinction between American English /r/ and /l/ and na-tive speakers of Japanese have a hard time discriminating between these two sounds. Here, we ask whether standard AutomaticSpeech Recognition (ASR) systems trained on large corpora of continuous speech can make correct quantitative predictionsregarding such non-native phonetic category perception effects. By training an ASR system on language L1 and evaluatingit on language L2, we obtain predictions for a native L1 speaker tested on L2 phonetic contrasts. Using a variety of L1 andL2, we show that ASR models correctly predict several well-documented effects. Beyond the immediate results, our evaluationmethodology, based on a machine version of ABX discrimination tasks, opens the possibility of a more systematic investigationof computational models of phonetic category perception.

How does a naive language learner deal with speaker variationirrelevant to distinguishing word meanings? Experimental datais contradictory, and incompatible models have been proposed.Here, we examine basic assumptions regarding the acousticsignal the learner deals with: Is speaker variability a hurdle indiscriminating sounds or can it easily be ignored? To this end,we summarize existing infant data. We then present machine-based discriminability scores of sound pairs obtained withoutany language knowledge. Our results show that speaker vari-ability decreases sound contrast discriminability, and that somecontrasts are affected more than others. However, chance per-formance is rare; most contrasts remain discriminable in theface of speaker variation. We take our results to mean thatspeaker variation is not a uniform hurdle to discriminatingsound contrasts, and careful examination is necessary whenplanning and interpreting studies testing whether and to whatextent infants (and adults) are sensitive to speaker differences.

This paper explores the minimal knowledge a listener needs tocompensate for phonological assimilation, one kind of phono-logical process responsible for variation in speech. We usedstandard automatic speech recognition models to represent En-glish and French listeners. We found that, first, some typesof models show language-specific assimilation patterns com-parable to those shown by human listeners. Like English lis-teners, when trained on English, the models compensate morefor place assimilation than for voicing assimilation, and likeFrench listeners, the models show the opposite pattern whentrained on French. Second, the models which best predict thehuman pattern use contextually-sensitive acoustic models andlanguage models, which capture allophony and phonotactics,but do not make use of higher-level knowledge of a lexiconor word boundaries. Finally, some models overcompensate forassimilation, showing a (super-human) ability to recover theunderlying form even in the absence of the triggering phono-logical context, pointing to an incomplete neutralization notexploited by human listeners.

Experimental research suggests that at birth infants can dis-criminate two languages if they belong to different rhythmicclasses, and by 4 months of age they can discriminate two lan-guages within the same class provided they have been previ-ously exposed to at least one of them. In this paper, we presenta novel application of speech technology tools to model lan-guage discrimination, which may help to understand how in-fants achieve high performance on this task. By combininga Gaussian Mixture Model of the acoustic space and low-dimensional representations of novel utterances with a modelof a habituation paradigm, we show that brief exposure toFrench does not allow to discriminate between two previouslyunheard languages with similar phonological properties, butfacilitates discrimination of two phonologically distant lan-guages. The implications of these findings are discussed.