Statistical learning (SL) is a powerful cognitive tool to extract statistical regularities from sensory input, which enables learners to detect structure in the vast amounts of sensory information they are exposed to. This sensitivity to statistical regularities has been investigated extensively in the field of cognitive science (Bogaerts, Frost, & Christiansen, 2020) as well as in linguistics given the important role it is postulated to play in the process of language acquisition (Romberg and Saffran, 2010, Saffran, 2003).

Two types of input regularities can be distinguished for SL (Siegelman et al., 2017, Siegelman et al., 2017). The first type pertains to sequential and spatial relations between stimuli, such as the co-occurrence of particular sounds and shapes in time (e.g., Saffran et al., 1996, Turk-Browne et al., 2005) or in space (e.g., Fiser and Aslin, 2001, Orbán et al., 2008). Because this type of SL refers to the learning of regularities such as conditional probabilities, we refer to it as conditional statistical learning (CSL, see also Thiessen and Erickson, 2013, Thiessen et al., 2013, Growns et al., 2020). As CSL helps to identify how sequences or complex scenes are formed from a set of discrete building blocks, it is often investigated in the context of speech segmentation and syntactic processing (Conway et al., 2010, Misyak and Christiansen, 2012) as well as in the context of visual processing (Fiser & Lengyel, 2022). The second type of regularity pertains to the frequency distribution of individual exemplars. Distributional statistical learning (DSL) has mainly been investigated in the context of phonology and category learning (Maye et al., 2002, Maye et al., 2008).

CSL and DSL are proposed to be supported by distinct yet interrelated memory processes and differ in the (implicit) knowledge acquired (Thiessen and Erickson, 2013, Thiessen et al., 2013). When facing a continuous sequential input, CSL relies on transitional probabilities to index sequential relations within the input stream and extract novel discrete representations of repeating patterns (Thiessen et al., 2013, Thiessen and Erickson, 2013). For example, the seminal study by Saffran and colleagues (1996) revealed that infants as young as 8 months old can track transitional probability information from an artificial speech stream of auditory syllables, allowing them to extract repeated words without any additional cues for word boundaries and to showcase familiarity with these words. In similar experiments as Saffran et al. (1996), this type of SL was found to facilitate the learning of various linguistic structures, including word order (Gervain, Macagno, Cogoi, Peña, & Mehler, 2008), syntactic patterns (Gomez & Gerken, 1999), and phonotactics (Chambers, Onishi, & Fisher, 2003). Outside the linguistic domain, the sensitivity to conditional relationships in sequential input was also demonstrated with visual (Kirkham et al., 2002, Zimmerer et al., 2010) and tactile stimuli (Conway and Christiansen, 2005, Conway and Christiansen, 2006). In the context of spatial input, sensitivity to conditional regularities was identified with visual stimuli featuring spatial configurations where multiple elements were presented simultaneously (e.g., Fiser and Aslin, 2001, Orbán et al., 2008). Moreover, CSL was documented across a wide range of age groups (Raviv and Arnon, 2018, Saffran et al., 1999) and even in non-human species (Milne et al., 2018, Sonnweber et al., 2015). Taken together, these results show that the sensitivity to conditional regularities in the input is present not only in human language learning but also in other domains.

By contrast, information regarding the frequency, variance, and context of multiple exemplars is aggregated during DSL. Thiessen et al., 2013, Thiessen and Erickson, 2013 proposed that, unlike CSL, where the extracted patterns form discrete representations in long-term memory, it is the integration across such discrete representations that gives rise to learners’ sensitivity to the distribution underlying the input and the discovery of categorical structure. For instance, Maye and colleagues (2002) showed that infants can categorize speech sounds taken from a phonetic continuum according to the bimodal frequency distribution of the input. After a familiarization phase with an input stream that either contained a bimodal or a unimodal frequency distribution of tokens from a [ta] to [da] continuum, only infants who were exposed to the bimodal distribution successfully discriminated tokens from the endpoints of the continuum. This result indicated that infants use distributional information to make sense of the acoustic variability that characterizes speech and learn the underlying phonetic structure of the language. Extending the results of Maye, Werker, and Gerken (2002), learning of the statistical distribution of sounds for forming phoneme categories was also documented in children (Vandermosten, Wouters, Ghesquière, & Golestani, 2019; see Cristia, 2018 for a review), adults (Hayes-Harb, 2007, Maye and Gerken, 2011), and non-human species (Pons, 2006). Considering other domains and types of input, DSL has been demonstrated with discrimination tasks of non-native lexical tones (Liu et al., 2022), musical pitches (Ong, Burnham, & Stevens, 2017), as well as shapes that differ in size (Rosenthal, Fusi, & Hochstein, 2001) and human faces (Altvater-Mackensen, Jessen, & Grossmann, 2017). In addition, recent investigations on the effect of statistical regularities on visual selection have found that participants give attentional priority to locations in a visual display where targets are likely to appear and suppress locations where distractors appear with higher probability (Theeuwes, Bogaerts, & van Moorselaar, 2022). This finding shows that DSL can also contribute to optimizing attention allocation and visual processing.

Although distributional patterns are a major component of language learning and general pattern detection, they have received much less attention. This is clearly illustrated in the review by Frost and colleagues (2019), which pointed out that the vast majority of SL studies have focused on sequential conditional regularities, using paradigms with embedded triplets and pairs akin to Saffran et al. (1996)‘s seminal study, or artificial grammar learning (Reber, 1969) which is commonly used in the implicit learning literature yet arguably measures a similar type of learning (Perruchet and Pacton, 2006, Christiansen, 2019). The focus of the current investigation is the modality- and stimulus-sensitivity of learning distributional regularities, and more specifically the learning of categories based on signals that vary in their length. In what follows we discuss previous research on the constraints on statistical learning at large, including CSL, as most works focused on the learning of sequential regularities and these findings set the stage for the current study and shaped our predictions.