Within linguistics and the cognitive sciences, some theories rely heavily on representations (Markman & Dietrich Reference Markman and Dietrich2000). As generic constructs, representations may be characterized as memory traces (Rosen Reference Rosen1975), categories (Rosch Reference Rosch1973), or concepts (Medin Reference Medin1989), depending on the level of abstraction pursued in the respective frameworks. Common to most approaches is the assumption that representations “stand for entities” in the environment (Ramsey Reference Ramsey2010), that is, that they provide essential units for simulating the external world. As such, representations are readily suited for predictive coding frameworks (Gładziejewski Reference Gładziejewski2015) that postulate that cognition is based on simulation, active inference, and hypothesis testing (Friston Reference Friston2005). In this vein, Gilead et al. propose that predictive cognition necessitates representations that are hierarchically organized along an abstraction continuum. The advantage of their approach is that it operationalizes the definition of abstraction, is readily applicable to phonetic and phonological speech sound representations, and is compatible with a range of neurolinguistic findings.

Several theories on sound representations in human language propose different levels of abstraction. Rooted in generative linguistics, a basic distinction is made between (more abstract) underlying representations and (less abstract) surface representations (Chomsky & Halle Reference Chomsky and Halle1968). This distinction can also be aligned with the distinction between phonology and phonetics. A phonological representation, in these approaches, is a phonemic (or meaning-distinguishing) representation of a sound in minimal pairs such as sum and sun. A phonetic representation, on the other hand, may be more detailed and provide additional information about pronunciation variants of the two word-final nasal consonants [m] and [n]. Abstraction is operationalized in approaches that consider speech sound representations to be a bundle of phonological or phonetic features (Clements Reference Clements1985; Jakobson et al. Reference Jakobson, Fant and Halle1965; Stevens Reference Stevens2002). The notion of abstractness in these theories closely resembles the definition provided by Gilead et al., in that abstractness means “two or more subjectively distinguishable objects [that] satisfy” the structural descriptions of the proposed representation (sect. 1, para. 1, first item 3). Distinguishable objects in speech sounds refer to acoustic–phonetic dimensions such as the place of articulation of nasal consonants, distinguishing bilabial [m] (produced at the lips) from alveolar [n] (produced at the alveolar ridge behind the upper front teeth). In English, on a relatively high level of abstraction, it is assumed that the nasal in morphemic in– (“not,” preceding probable or tolerant) shares the “subjectively [and phonetically] distinguishable objects” bilabial and alveolar place of articulation to allow for the variations seen in improbable and intolerant. Because both places of articulation cannot be expressed simultaneously, the assumption in phonological theories is that the abstract representation of nasal consonants does not include place of articulation information in such cases. This is referred to as underspecification (Lahiri & Reetz Reference Lahiri and Reetz2010; Steriade Reference Steriade and Goldsmith1995): The nasal consonant in improbable and intolerant is said to be underspecified for place of articulation. In the process of speech sound production, the spell-out of more concrete representations (concreta in the terminology of Gilead et al.) is accomplished by phonological and phonetic rules (criterion of substitutability in the terminology of Gilead et al.). These rules insert the necessary place of articulation feature either as a result of a default setting or as a result of a contextual fill-in (Lahiri & Reetz Reference Lahiri and Reetz2010). Similar to Gilead et al.'s approach, the outputs of abstraction processes can serve as inputs for additional abstraction processes. The feature bundle describing [m] would contain the feature consonantal (distinguishing consonants from vowels), nasal (distinguishing nasal and non-nasal consonants), and labial (distinguishing labial and non-labial consonants). The additional act of abstraction in instantiating underspecification would then remove the feature labial.

Abstract (underspecified) speech sound representations may relate to predictive processes in at least two ways. First, the spell-out of concreta (phonetic representations) can be context-specific in that following consonants determine the place of articulation of the preceding nasal consonant, leading to [m] when preceding bilabial [p] (improbable) and to [n] when preceding alveolar [t] (intolerant). Gilead et al. propose that “there is no mental travel without abstraction” (sect. 4, para. 2) where mental travel stands for predicting future states; in the speech example, predicting future states would mean predicting upcoming articulations and modifying implementations of abstract representations accordingly. During speech production, simulations of upcoming articulations account for the observation that upcoming speech sounds indeed exert influences on earlier speech sounds. This regressive assimilation has been related to anticipatory processing (Gow Reference Gow2001). Internal simulations of future states may also minimize the prediction error in the production–perception loop involved in speech processing. Second, speech sounds differing in their assumed degree of abstraction seem to relate to differences in predictive processing. Broadly speaking, speech sounds based on more features allow for more precise or stronger predictions. Evidence for this observation stems from neurolinguistic studies (e.g., Scharinger et al. Reference Scharinger, Monahan and Idsardi2016) interpreted within an auditory predictive coding framework (Baldeweg Reference Baldeweg2006). Further evidence is provided by studies that suggest more concrete representations to be based on supplementary motor information (Möttönen & Watkins Reference Möttönen and Watkins2009; Möttönen et al. Reference Möttönen, Dutton and Watkins2013). For instance, by disengaging the motor lip area, Möttönen et al. (Reference Möttönen, Dutton and Watkins2013) showed that representations of labial speech sounds (in ba-syllables) were disrupted, whereas representations of non-labial speech sounds (in da- or ga-syllables) were not.

Abstraction in speech sound representations as illustrated above is reflected in cortical structures in human temporal cortices where increasingly abstract speech sound representations are supported by regions with increasing distance to the primary auditory cortex (Humphries et al. Reference Humphries, Sabri, Lewis and Liebenthal2014; Obleser & Eisner Reference Obleser and Eisner2009). In these studies, more detailed and concrete (phonetic) representations were supported by regions in primary auditory cortex in the vicinity of Heschl's gyrus, whereas more abstract (phonological) representations recruited regions in more anterior and posterior parts of the superior temporal gyrus and sulcus. This cortical abstraction gradient is parallel to the one for visual perception by Gilead et al. and is in line with the aforementioned phonetic-phonology distinction.

In sum, Gilead et al.'s approach provides interesting starting points for future studies combining current theories of brain function with generative linguistic accounts of speech sound representations.