2.1 Spell-out

Introduced by Hale & Reiss (2000), Substance-Free Phonology is just what its name says it is: the idea that phonology operates without any reference to phonetics, and without knowing that the units manipulated will at some point be pronounced and have a phonetic existence. Thus phonological primes are units whose only purposes are

1. to express contrast

2. to allow for phonological computation

It follows that no name can be given to phonological primes since their only identity is to be distinct from others. Greek letter symbols α,β,γ used to refer to them are just a crutch for our convenience, arbitrarily chosen and interchangeable: whether alpha names the phonological item whose phonetic correlate is labiality (α ↔ labiality), or beta (β ↔ labiality), or banana (🍌 ↔ labiality), is irrelevant.³

Phonological primes are eventually pronounced and therefore have a phonetic correlate. In production, phonetic categories come into being when the output of phonology is transmitted to phonetics. This implies a modular setup whereby phonology and phonetics are ontologically distinct computational systems working on distinct vocabulary (Fodor 1983 followed by, among many others, Coltheart 1999; Gerrans 2002; Jackendoff 2002; Carruthers 2006: 3ff; Reiss 2007; Boeckx 2010; Robbins 2017).

The transmission is effected through a look-up table containing a list of spell-out instructions where phonological items are matched with phonetic items, just like in a language dictionary (where e.g. English dog is matched with French chien). This is shown under (4) below. Note that the symbol used for matching items of distinct vocabularies is ↔ in post-phonological spell-out, purposefully the same that is used when morpho-syntactic items are translated into phonological items upon post-syntactic spell-out, which is just as arbitrary and involves the exact same kind of look-up table (e.g. in English, past tense ↔ -ed). That is, intermodular communication uses only one mechanism, whatever the modules (Scheer 2014; 2020; 2022).

This list-based translation appears in various guises in the literature: as cue constraints in the BiPhon model (Boersma & Hamann 2008; Hamann 2011), as phonetic interpretation (Harris & Lindsey 1995: 46ff; Gussmann 2007: 25ff; Cyran 2014), as transduction in Hale & Reiss (2008: 109–118), or in its incarnation used here, post-phonological spell-out (Scheer 2014; 2019; 2022).⁴

The phonetic correlates under (4) are articulatory properties for expository reasons. In the BiPhon model (Boersma & Hamann 2008), the vocabulary of the phonetic computational system is acoustic: the auditory continuum is expressed as an auditory spectral mean along the ERB scale (Equivalent Rectangular Bandwidth). That is, spell-out associates for example s ↔ 20,1 ERB. This reads “s is realized with a spectral mean of 20,1 ERB”. The discrete ERB value thus defines the acoustic target of the pronunciation to be realized. The computation carried out in the phonetic module then creates slight variation around this target.

In the discussion below, we will meet the idea that production involves an acoustic target in the guise of the structuralist notion Lautvorstellung “imagination of a sound” (section 3.5).

2.2 Computation

Given spell-out, a regular palatalization expressed with classical substance-laden primes under (5a) identifies as (5b) in the substance-free setting, where α, β and γ have the phonetic correlates under (4).

Note that phonological computation is arbitrary: it may add or remove primes (associate or dissociate in an autosegmental structure) in any way, and the trigger may be anything or its reverse. Thus any segment may be turned into any other segment in any context (Hale & Reiss 2008: 162–171). Danesi (2022) discusses in detail how primes and their phonetic correlates are established based on the alternations found in a language, and how substance-free computation works given these settings.

2.3 There is nothing wrong with being lazy

In practice, there is nothing wrong with using substance-laden labels when talking about phonology, provided it is understood that this is merely shorthand for (5b). Chemists talk about water when they mean H₂O, and black holes are neither black nor holes, but astronomers are happy to use this term. Phonologists also talk about, say, cat in various degrees of explicitness: cat in spelling, [kæt] in IPA; syllable structure may be added to that, as well as the featural makeup of segments.

It may thus be understood that substance-laden features are in fact alphas, betas and gammas, but that it is more convenient to talk about their phonetic correlates when referring to them. In fact much of the SFP literature does exactly that (sometimes without making laziness explicit, though). That is, phonologists are merely lazy, trying to facilitate communication when using substance-laden labels to talk about phonology, just as chemists are lazy when talking about water. The difference is that every chemist knows that water is not the real object they are talking about, while phonologists are not aware of their laziness: they believe that the phonological item they talk about really has phonetic properties. One goal of SFP is to raise phonologists’ awareness that they are lazy.

2.4 No naturalness (in phonology)

Phonology has no clue how its product will eventually be pronounced, or indeed that it will be pronounced at all: this follows from the fact that phonological primes α, β, γ are substance-free. A direct and necessary consequence is that no phonetic constraint can bear on phonology. This means, in turn, that naturalness has no business in phonology. On the question of naturalness in the phonological literature and its status in SFP, see Chabot (2021).

The mainstream view going back to the 19^th century is that phonology is responsible for naturalness, i.e. the observation that the overwhelming majority of phonological patterns and processes make sense phonetically speaking. Thus a palatalization of the kind shown under (5a) (k,g → ʧ,ʤ / __i,e) will be called a natural process, while English velar softening where the output of palatalization is s (k → s / __i as in electri[k] – electri[s]-ity) will be regarded as suspicious, also because it does not occur morpheme-internally (king is not pronounced [sɪŋ]) and is only triggered by class 1 suffixes (Kiparsky 1982: 40f). Finally, there are processes that are called crazy rules (Bach & Harms 1972) because they make no phonetic sense: i → u / d__ found in Kashaya Pomoan (Pomo, California) is a case in point (Buckley 2000; 2004).

The only possible definition of naturalness is phonetic: something is natural because the phonetic properties of an item influence the phonetic properties of another item (palatalization is assimilation: front vowels make non-front velars more front).

If there is no phonetics in phonology, there is no way phonology could possibly be held accountable for naturalness, or its absence. If phonology is made of α, β, γ, it is not even possible to talk of anything being natural or unnatural. Thus the substance-free statement (5b) β, γ → β, γ + α / __ α may refer to the perfectly natural palatalization (5a), but could as well describe the reverse unnatural event ʧ,ʤ → k,g / __i,e, or the crazy Kashaya Pomoan rule. The SFP literature discusses a number of formal (i.e. substance-unrelated) properties that restrict what computation can do,⁵ but these do not bear on the fact that any segment may be turned into any other segment in any context. Therefore phonology is just as happy to compute natural rules as it is to compute suspicious or crazy rules.

Naturalness in sound patterns surely exists, but has nothing to do with phonology. Its only source is phonetics, and it arises because phonological processes have phonetic precursors, which at some point are phonologized. Phonetic precursors are always natural since they are phonetic, and phonology inherits them (e.g. Bermúdez-Otero 2015).

In section 3.6 below, we will come across an example of naturalness, chain shifts restoring the equilibrium of vowels in phonetic space, which structuralist analysis takes to be a consequence of a phonological reaction against an imbalanced system. This phenomenon also has a phonetic rationale, vowel dispersion, which by SFP standards is the only viable explanation.

2.5 Two schools in SFP: primes are emergent or universal and innate

All versions of SFP agree on the workings of spell-out described in the previous section. There are two competing views on other properties of spell-out, though. Authors working (or formerly working) at Concordia University in Montreal, i.e., where Hale and Reiss found their academic home, hold that both primes and their association to phonetic categories are genetically coded and present at birth: there is a universal set of around 20–30 substance-free primes (α, β, γ) from which L1 learners draw items when building their system, and which are associated to specific phonetic categories (Hale et al. 2007: 647f; Volenec & Reiss 2018; see also Hale & Reiss 2000; 2008; Reiss 2017; Reiss & Volenec 2022).

All other representatives of SFP, declared as such (Blaho 2008; Odden 2006; 2022; Samuels 2011a,b; 2012; Scheer 2014; 2022; Iosad 2012: 6ff; 2017; Chabot 2019; 2021; 2022; 2023; Danesi 2022) or under a different heading (Boersma 1998: 461ff; 2011; Boersma & Hamann 2008; Hamann 2011; 2014; Mielke 2008; Carvalho 2002; Dresher 2014; 2018), hold that primes are emergent, i.e. absent at birth, and that their association with phonetic categories is learned during first language acquisition. Overviews of this landscape are proposed by Scheer (2019) and Samuels et al. (2022).⁶

2.6 Arbitrariness of association

In the emergentist view, the association of a phonological prime with a phonetic correlate is arbitrary. Any phonetic category may in principle be associated to any prime: there are no cross-linguistically stable or innate associations. The symbol ↔ materializes the arbitrary character of the association just like at the upper interface of phonology, where any morpho-syntactic property may be associated to any phonological item: there is no cross-linguistically stable or innate association at this interface either.

At both interfaces, the associations are language-specific, learned during L1 acquisition, stored in long-term memory and subject to diachronic evolution. That is, the association past tense ↔ -ed is in place in English, but arbitrary in the sense that in another language past tense may be associated to another phonological item. The same holds for spell-out instructions at the phonology-phonetics interface, e.g. for α ↔ labiality.

3.1 Kruszewski

Baudouin de Courtenay (1895: 7, note 1) acknowledges that the distinction between sound (German Laut) and phoneme (German Phonem) is due to his student Mikołaj Kruszewski (both being Polish, they worked at the University of Kazan in Russia, and are referred to as the Kazan school, see Radwańska-Williams 2021; Anderson 2021: 69ff). He adds that his understanding of the phoneme has changed over time, originally leaning on the one introduced by Kruszewski (1881) when he coined the word phoneme.

Kruszewski defines the phoneme as follows, in a few words only and without further elaboration:

Kruszewski‘s idea is to separate what humans produce (sound, referred to as the anthropophonic unit) and what the phonetic identity of these sounds is: this identity is phonetically atomic, i.e. cannot be further divided. The idea that the variation produced by pronunciation is the expression of a single unit, the phoneme, will have an important offspring (see section 3.5).

But for now let us note that the phoneme enters the scene in an explicitly phonetic guise: it is entirely made of phonetic substance.⁸

3.2 Trubetzkoy

Representatives of the Prague school leave no doubt that phonological units do bear phonetic content. This is because, they argue, phonemes are identified by their phonetic properties. How would you know which phoneme is which other than by their phonetic characteristics?

In his Grundzüge, Trubetzkoy is explicit on that: in the quote under (7), the distinctiveness of the phonemes at hand is based on the phonetic property of voicing which, Trubetzkoy says, must therefore characterize phonemes in the phonology.

Trubetzkoy tries to have his cake and eat it too, though: yes phonological units have phonetic content, but sure phonology and phonetics are fundamentally independent. It is just the word fundamental and its interpretation that saves Trubetzkoy from schizophrenia. This is the consciousness of sin tempting Trubetzkoy and others into a contradiction: there is no explanation how phonology could be independent of phonetics, while accommodating it. Being thus conscious of his sin, Trubetzkoy continues the quote under (7) by saying that maybe there is phonetics in phonology, but just a little, only what is strictly necessary. Note that the meaning of the crucial word in the quote below, Anfangsteile in the original German text, is entirely obscure (literal translation: beginning (or initial) pieces).⁹

What Trubetzkoy says expresses the doctrinal body of the Prague School defined in the nine theses published (unattributed) in the first issue of the Travaux du Cercle Linguistique de Prague. Here is a piece of thesis Nb. 2:

Thus the phonetic content of a phoneme is «less essential» than the correlations it contracts with other phonemes. But it exists. Also note that the set of phonemes of a language is defined as the list of «acoustico-motor images» (on which more in section 3.5).

3.3 Jakobson

The quotes under (2) and (3) have shown Jakobson’s (1949) consciousness of sin against the Saussurian insight. The latter quote ends with a question mark: when we determine the “specific essence” of distinctive features, are we sinning against de Saussure? The answer appears on the following page: yes we are, and there is no way to avoid the sin.

Note that “the phonic nucleus” (where phonic means phonetic) of an opposition is its nucleus, i.e. its most intimate property. This is not being lazy: phonetics are really an essential property of the phoneme, says Jakobson.

Trubetzkoy’s and Jakobson’s argument for the presence of phonetic content in phonemes is thus that in its absence, it would be impossible to know which one is which phoneme: “the phonic nucleus” is “utilized semiotically”, i.e. to distinguish phonemes.

This argument is sound, but Trubetzkoy’s and Jakobson’s conclusion that therefore the pronunciation must be in-built into the phoneme is faulty: spell-out was not on their radar. The association of phonological and phonetic items as under (4) in a spell-out dictionary allows phonologists to recant sin. Spell-out identifies phonological items by how they are pronounced, yet without placing the pronunciation in the phonology. Jakobson’s (and Trubetzkoy’s) redemption hinged on spell-out, which makes good on de Saussure’s promise.

3.4 Bloomfield

Leonard Bloomfield describes how those acoustic properties that are distinctive are bundled by the listener and in their acoustic shape constitute what is called a phoneme:

While distinctive features were introduced in 1928 by Jakobson et al. (1928), they did not run under this name then (see note 3). Bloomfield uses the term distinctive feature already in 1933, and says they are the building blocks of phonemes. There is no doubt that Bloomfield really means that acoustic properties are present in the phoneme: acoustic features are “recurring in recognizable and relatively constant shape in successive utterances”. They are thus portions of the acoustic signal. In production, Bloomfield says, the speaker has learned to articulate in such a way that “phoneme-features will be present in the sound-waves”.

3.5 Sin in the definition of the phoneme: Lautvorstellung

Twaddell (1935) reviews different ways in which the phoneme is defined in the literature of his time (see Anderson 2021: 376ff). He classifies the approaches into two groups: those contending that the phoneme is a mental or psychological item (what today would be called cognitive), and those arguing that it is a physical reality.

The latter goes back to the definition by Kruszewski (1881) (recall section 3.1). The relevant observation is that speech sounds have infinitely variable pronunciations, but still represent the same object: the p in path has millions of different phonetic realizations, but still is a “p” for both the speakers and the perceivers. “It is the recognition of this sameness, this effective unity, which has found expression in the term ‘phoneme’ as a unit of spoken language” (Twaddell 1935: 5). Twaddell further explains that the phoneme is a creature of the increased instrumental accuracy of phonetic investigation:

The competing mentalist idea is well described by the German word Lautvorstellung “the imagination of a sound” found for example in Baudouin de Courtenay (1929: 39) or Trubetzkoy (1929), who uses Phonem and Lautvorstellung synonymously:

Lautvorstellung also appears as Psychophon (as opposed to Physiophon, Benni 1929: 35f), image articulatoire et auditive “articulatory and auditive image” (Sommerfelt 1971 [1928]: 199) or as images acoustico-motrices “acoustico-motor images” in the definition of the doctrinal body of the Cercle Linguistique de Prague in 1929, see the quote under (9). Twaddell (1935: 8f) quotes still more authors going by this definition of the phoneme, and interestingly not a single source is written in English.

What this means is that the cognitive unit phoneme consists of a phonetic target: the auditory and articulatory image of the pronunciation that was intended by the speaker. This intended pronunciation then experiences a number of distortions when put to use, creating the phonetic variation observed. The task of the listener is to reconstruct what was intended from the variable acoustic signal.¹⁰

The idea that a phonological unit is externalized thanks to an acoustic target is taken up in much modern work, including the literature discussed in section 2.1. It is an exact instantiation of the spell-out mechanism (4), except that the acoustic target was misplaced by Trubetzkoy and colleagues: since they had not thought of a spell-out mechanism, the only possible locus of the acoustic target was in the phoneme itself. That is, the phoneme is the Lautvorstellung = the acoustic target. With a spell-out mechanism, the phonological unit (phoneme or feature) is exempt from any acoustic target. Rather, the acoustic target is the phonetic item into which the phonological unit is translated (through a look-up list).

The conclusion is the leitmotif of this article: most structuralists, here those of the Prague School, had to live in sin because they had not thought of phonology and phonetics in terms of a modular structure where both systems are related by a spell-out mechanism.

3.6 Chain shifts

The basic building blocks of structuralist analysis are phonemes. A substantial concern of structuralism is how exactly linguists can determine the set of phonemes and their relationship (correlations), given the phonetic surface. This was called the discovery procedure, or phonemic analysis (e.g. Hockett 1942: 20f; Dresher 2009: 13–19). But the literature is less loquacious explaining how a phoneme becomes a phonetic realization upon speech production.

Structuralism is known for the large amount of slack that was allowed between phonemes and their phonetic realization. Or rather, for the lack of ambition to determine which allophone deserves phonemic status. Thus in a language where [e] and [ɛ] are in complementary distribution, will the phoneme be /e/ or /ɛ/? Structuralists would not bother with this question and rather say that the phoneme is /E/, i.e. a vowel located in the area of /e/ and /ɛ/. It is not useful to be more precise regarding its phonetic properties since apart from its correlations with other units, the only important thing for a phoneme is to be different from other phonemes, here from /O/, /I/, /U/, /A/ in a 5 vowel system.

How much [e] or [ɛ], if any, is there in a structuralist phoneme referred to as /E/? Is E chosen to substantiate the phoneme at hand because of its phonetic essence, a front mid unrounded vowel? Or are the phonetic properties used to characterize phonemes just shorthand for naming phonological items that do not have any phonetic properties? In other words, is this merely a case of laziness in the sense of section 2.3? The answer is no: structuralist phonemes are located in the vocalic triangle and thus have a phonetic anchor. By contrast, alphas and betas have no location in any phonetic space. There is no way in which α is any more front, back, round or high than β etc. Alphas and betas as well as the phonological computation that manipulates them have no clue that phonetic substance exists, and that they will eventually end up being pronounced.

de Saussure’s perspective, again, is fully substance-free. Given the quote under (14), he disagrees with the presence of any phonetic content in phonemes. The only thing that makes the identity of a phoneme, its only property, he says, is its distinctiveness with respect to other phonemes.

A central idea that was introduced by structuralism and makes a good deal of its identity is that phonemes are arranged in systems, which produce systemic pressure: they strive for equilibrium. Chain shifts (Gordon 2011) such as the Great Vowel Shift that occurred at the end of the Middle English period (e.g. Krug 2017) are typical illustrations of systemic action: given a symmetrical system in equilibrium (15a), the initial spark is a change in the pronunciation of a phoneme or of a set of phonemes caused by some phonetic or other factor, here the diphthongization of high vowels under (15b). Once the phonetic space is thus vacated, neighbouring phonemes move to occupy it: under (15c), E and O raise to I and U, respectively. That is, they change their pronunciation because they are phonetic neighbours: Martinet (1955) calls this movement a drag chain or pull chain (see Anderson 2021: 218ff).

On the structuralist analysis, chain shifts occur in the phonology: the items involved are phonemes, and the evolution observed is caused by a systemic reaction. The phoneme labels necessarily note phonetic content that is present in the phonology: E and O need to “know” that they are mid and therefore move in order to fill in the vacancy left behind by their phonetic neighbours I and U. A does not move into the position of I or U because it is not phonetically adjacent.

Also note that on the structuralist analysis, phonemes change labels when their pronunciation is changed: E raises to occupy blank phonetic space, and as a result becomes I. This also shows that the phonological identity of phonemes is determined by the way they are pronounced.

In a substance-free perspective where the phonological identity of phonemes has nothing to do with how they are pronounced, the chain shift under (15) involves five phonemes, whose pronunciation changes over time, as under (16). Their labels cannot change because they do not have any: a given phoneme, say β under (16), is pronounced [e] before, [i] after the shift.

The shift is caused by the same mechanism that was identified by the structuralists, but has nothing to do with phonology or phonemes. The description of the facts refers only to phonetic items: how phonemes are pronounced, how a blank spot is created in phonetic space, and how the pronunciation of other phonemes changes in order to rebalance the distribution of phoneme pronunciations in the phonetic space.

The number of phonemes and their correlations have not changed over time: the only thing that has changed is their pronunciation, creating an unequally populated phonetic space.

The structuralist idea of self-equilibrating systems was based on the observation that given the articulatory (and acoustic) space defined by buccal anatomy, vowel and consonant systems place individual vowels/consonants in such a way that they equally occupy this space. That is, the distance between vowels (or consonants) is maximized and their distribution is symmetric. A number of asymmetric systems notwithstanding, this is confirmed by modern typological reviews, e.g. Lass (1984), Clements (2003; 2009), Hall (2011). Following the structuralist model, these authors devise phonology-internal mechanisms to account for the observation.

An extra-phonological alternative is dispersion, a phonetic rationale based on perception, learnability or other non-phonological mechanisms. The most popular perception-based approach holds that optimizing the articulatory space by maximizing distance and phonetic dissimilarity among segments enhances perceptibility (Liljencrants & Lindblom 1972; Lindblom 1986; Schwartz et al. 1997; Johnson 2000). Vaux & Samuels (2015: 573) argue that dispersion is due learnability: “[a] system that is easier to learn (e.g., one whose members are more dispersed in perceptual space) is predicted by our model to become more common cross-linguistically over evolutionary time than its less learnable competitors.”

Grammar or UG have nothing to do with the self-equilibration of systems via dispersion: phonology is just as happy with a dispersed i-a-u system as it would be with an undispersed i-e-a or y-ə-o system. If phonology is substance-free, the self-equilibrating virtue of segment inventories in the phonetic space has nothing to do with phonology or grammar: it is a purely phonetic event due to functional factors.

This section was to show that typical structuralist labels of phonemes such as E are not just shorthand. That is, the label E is not chosen out of laziness, but rather describes what is held to be the true phonetic content of the phoneme. Thus on this common structuralist take, phonetic content is present in the phonology.

4.1 Baudouin de Courtenay

There are also structuralists who deserve grace: Baudouin de Courtenay and Louis Hjelmslev (see chapters four and seven of Anderson 2021 for background information).

At the outset of his book on alternations, Baudouin de Courtenay (1895) provides some “explanation and definition of terms”, among which the phoneme:

Thus the phoneme is the “psychological equivalent of the speech sound”, built in the soul upon exposure to phonetic input. The phoneme is associated with articulatory and acoustic values. This is an exact description of the workings of spell-out under (4): the phonological item is associated with phonetic correlates, but has no phonetic content itself.¹¹ Baudouin de Courtenay does not dwell on how exactly the association converts phonological into phonetic items, though.

4.2 Hjelmslev

6.1 Generative radicalization

After having discussed the generative position on the presence of phonetics in phonology in the previous section, let us now look at the broader context of the exchange with structuralism. Starting with Chomsky’s (1951) Masters thesis, generative phonology was built during the 1950s and 1960s (Anderson 2021: 403ff). It faced what Postal (1968: ix) calls the entrenched dominant view in the United States. Generative phonology took issue with a number of tenets of the dominant theory, especially regarding two points: on the one hand the phoneme and more largely contrast (Dresher & Hall 2020), which were referred to as autonomous phonemics (or Chomsky’s 1964: 75ff taxonomic phonemics) by their generative contenders who denied their existence; on the other hand, the idea of rewrite rules (Hockett 1954; Halle 1962), which at that point were shared by syntax and phonology on the generative side (this is why the theory was called transformational).¹³

Both issues relate to the question of phonetics in phonology. Without the phoneme, phonology lacks units that can be defined independently of phonetic properties, that is as structuralist relations (here: contrast). The only alternative for the makeup of phonological units is then phonetic content. Rewrite rules operate in the phonology, where they transform a lexical item into a phonetic one. This can only work if the type of objects manipulated is the same: phonological items must be phonetic in kind since phonetics is the output of the transformation. A /p/ can be transformed into an [f], but there is no way to even talk about a putative transformation of an alpha into an [f].

The alternative is made perfectly explicit by contemporary Stratificational Grammar (Lamb 1966a) and Fudge (1967): if contrast is relevant, the units of phonology must be phonemes, which in turn must be bare of phonetic content and therefore identify as alphas and betas (Hjelmslev’s unnamed entities). From this it follows that the transition from phonology to phonetics cannot be rewrite rules: it makes no sense to transform an alpha into a [p]. This is why phonological units are related to phonetic items by spell-out, which appears as realization in Stratificational Grammar (an alpha is realized as [f]), and as articulatory rules in Fudge (1967). The appendix discusses Stratification Grammar (section A2.2) and Fudge (1967, section A2.4) in greater detail.

Representatives of the then-new generative view constantly appeal to Roman Jakobson as the godfather of phonetically imbued phonological items. Their view on the presence of phonetics in phonology is radicalized with respect to Jakobson and the Prague school, though, for the reasons mentioned: the absence of contrast and the phoneme, as well as the transformational take on the relationship of phonology with phonetics, which enforces an absolute reliance on the phonetic vector all through. This is reflected in Postal’s (1968) book entirely devoted to a rebuttal of autonomous phonemics in general and Stratificational Grammar in particular. Section A2.3 expands on this view.

Therefore, the parties opposing Louis Hjelmslev and the one that righteous phonologists face in generative times are quite different in kind. Hjelmslev could argue to be true to the structuralist reference de Saussure, who was respected by everyone around. And his opposing party was aware that they were sinning against the foundational statement of the theory that they held themselves to represent. This doctrinal ballast was removed from the shoulders of generativists who could wholeheartedly sin without being sinners. Thus the environment was arguably quite a bit rougher for a righteous phonologist in early generative times than it was before.

But still some upheld the Saussurian torch in the generative blizzard.¹⁴ In the appendix three cases are discussed: Voronkova & Steblin-Kamenskij (1975 [1970], see section A2.1), Erik Fudge (see section A2.4) and representatives of Stratificational Grammar such as Sidney Lamb and David G. Lockwood (see section A2.2). The vivid reaction of the generative party is a good measure of how crucial the presence of phonetics in phonology was for generative theory.

6.2 SPE considers and dismisses spell-out

Following the discussion with structuralist positions in the 1950s and 1960s mentioned in the previous section, the generative take on phonology was defined in The Sound Pattern of English (SPE, Chomsky & Halle 1968). In a section called Phonological and phonetic representation (p. 164ff), the authors introduce phonetic distinctive features and phonological distinctive features, which they say have unfortunately been confused since distinctive features were introduced.

Phonetic features are made of a phonetic label (called a phonetic scale) and an integer, the latter indicating the label’s strength for a given segment: thus /n/ may be [1 nasal] or [3 nasal] depending on context.

As shown by the quote, phonetic distinctive features are universal, drawn from a predetermined set. Interestingly, phonetic features are called distinctive, just like phonological features. This is in line with the generative dismissal of contrast (see section 6.1), which by structuralist standards exists only in the phonology: in its absence, “distinctive” refers broadly to anything that makes things different, here phonetic properties that have been chosen from the universal set of phonetic features.

An interesting question of course is how these phonetic properties are picked from the phonetic continuum. On p. 298, the authors say that anything that is “clearly an independently controllable aspect of the speech event” will be a phonetic feature: talking about labio-velar consonants in Yoruba that are produced with a “special clicklike suction”, they say that “[s]ince clicklike suction is clearly an independently controllable aspect of the speech event, the data just cited establish suction as a separate phonetic feature, regardless of the fact that apparently in no language are there contrasting pairs of utterances that differ solely in this feature.”

In a section called Phonetic and phonological representation (p. 295ff), Chomsky & Halle then ask what phonological features may look like, given the universal set of phonetic features. They conclude that the labels of phonological features are the same as those of the phonetic features.

This is the result of a discussion where the authors consider alphas and betas (here A, B) in an earlier chapter:

Then immediately preceding the decision (32), they also consider the spell-out that is needed for alphas and betas to be converted into phonetic items:

It is unclear why A, B, C, D, which clearly represent phonological features, are converted into whole words, rather than into segments, but that does not need to concern us here: what is relevant is that Chomsky & Halle consider spell-out based on alphas and betas, by what they call ad hoc rules.

The authors reject this option for four reasons. First, ad hoc rules are an unnecessary complication: given the universal set of phonetic features, all languages would have to have the same instructions. This would be redundant and could not help representing the difference between languages. Such redundancy can be avoided by directly giving phonological features the label of phonetic features:

This argument rests on the idea that there is a fixed universal set of phonetic features, which is not self-evident. Voronkova & Steblin-Kamenskij (1975 [1970]) argue against this idea, which is reminiscent of the position taken by the Concordia school of SFP: there is a set of universal and innate (phonological) features, whose phonetic correlates are also universal and innate. The alternative entertained elsewhere in SFP is that features and their phonetic correlates are emergent (see section 2.5).

The second argument is about language acquisition, which Chomsky & Halle (1968: 170) use to motivate identical phonetic and phonological features. This identity facilitates the task of the learner, they argue, who does not need to bother with (distributional) analysis in order to discover the phonological items that need to be stored in the lexicon: the child has the universal set of phonetic features at birth and merely needs to detect their phonetic values in the acoustic signal in order to know what is phonological and lexical.

The third argument is about what today are called natural classes:

The question is why we would like or need to express natural classes in the grammar. For Chomsky & Halle, the only possible origin of sound patterns lies in the phonology: they do not consider the logical possibility that they may also have phonetic causes, which are then only administered by the phonology (see section 2.4). Section A2.3 shows how Postal (1968) expands on this issue.

Finally, the fourth argument is shown under (37) below.

In my understanding (helped by the insights Elan Dresher shared when I asked him for advice), the argument is based on the fact that in SPE a feature matrix is made of two distinct types of features: phonetic features which co-define individual segments of a morpheme, and features like [+latinate] that define entire morphemes only, i.e. not individual segments. This is the formal property of language that Chomsky & Halle refer to which, they argue, would disappear if phonology were only made of alphas and betas: [+labial] would be [+A] and [+latinate] would be [+B]. The two would be indistinguishable. Chomsky & Halle do not exclude that individual segments may also be marked by [+latinate]-type features and then be targeted by computation (“rules of other sorts”). These will however add extra complexity to the grammar, which may be disfavoured by an evaluation measure, discussed in the following section (pp. 172–176) and whose precise contours is “far from definitive” (p. 175). The reasoning is thus what was called a potential argument at the time: one relying on premises that may not be correct, but in case they are produce the argument.

If this is the correct interpretation, the question arises in which way [+labial] and [+latinate] are better candidates for representing the formal property of language at hand: how are they distinct in kind? Are they any more intrinsically distinct than an alpha and a beta? Grammar must somehow know that [+labial] refers to a given segment, while [+latinate] is a property of the whole morpheme. The same would be true if phonological units were alphas and betas: grammar would need to know that alpha defines a segment, but beta the morpheme.

To conclude, Chomsky & Halle explicitly reject any kind of spell-out. The actual phonetic features occur in the lexical recording through all phonological operations into phonetics, where they become gradient, rather than binary.

Dresher & Hall (2020) discuss Halle’s Sound Pattern of Russian (Halle 1959) where features are organized into a contrastive hierarchy. This was abandoned in further generative development, as indeed contrast was. Dresher & Hall show that if Halle had used phonological processing for establishing the hierarchy, rather than the wish to use as little features as possible (simplicity), generative phonology may have evolved in a different direction where contrast would have been maintained and used to account for phonological processing. Dresher & Hall’s article is about this road not taken.

Taking over the Saussurian and Hjelmslevian view on the identity of phonological units is another road not taken, this time in SPE. Chomsky & Halle consider but dismiss alphas and spell-out, for reasons where simplicity also plays a role: to them, spell-out is redundant and an unnecessary complication (35); the task of the child is simplified when phonological and phonetic units are identical.

Clearly Chomsky & Halle (1968) were not concerned with what would later run under the label of modularity (Fodor 1983), although the recognition of separate computational systems using proprietary vocabulary was introduced in Chomsky’s Logical Structure of Linguistic Theory (LSLT; written in 1955–1956 but only published as Chomsky 1975) and put to use in the inverted T model in Aspects (Chomsky 1965) (more on that in section A2.2.2). Had they taken modularity seriously, there would have been no way to have the same vocabulary in both phonetics and phonology, and spell-out would not have been a redundant and unnecessary complication, but rather a crucial condition on phonetics and phonology being separate computational systems (just as spell-out, i.e. Vocabulary Insertion, at the syntax-phonology interface is, including in SPE: 163).

That a generative phonology with alphas and spell-out is possible has been shown by SFP since its inception. Representatives of Concordia in particular argue for a phonology that is close to SPE, except with alphas and spell-out. A price that Chomsky & Halle were not prepared to pay is the elimination of naturalness from phonology (36).

6.3 How do you measure abstractness when comparing alpha to beta, or alpha to [p]?

As was shown, the cornerstone of SPE’s approach to melodic primes is that they are identical i) in the lexicon, ii) in the phonology and iii) in the phonetics. This position is a direct consequence of rewrite rules, which were new in the field: they require a basic vocabulary that is manipulated by the rewrite computation whereby a lexical item is the input and a phonetic surface item the output. A trivial property of computation is that the input and the output belong to the same vocabulary. Phonemes (and therefore contrast) have no place in this scenario since they exist only in the phonology: there is no phonetic version of a phoneme, which therefore could not be rewritten as a phonetic object. Only spell-out based on a look-up table (section 2.1) can transform items of one vocabulary into items of another vocabulary. This is what Chomsky & Halle dismissed because they were operating in a non-modular setting where phonology and phonetics belong to the same computational system. It was mentioned that this is surprising given that modularity was a cornerstone of Chomsky’s LSLT (see section A2.2.2).

Locating phonology and phonetics in the same computational system allowed for measuring the distance between lexical and phonetic forms. This is the purpose of Kiparsky’s article How abstract is phonology? which eventually led to the emergence of Lexical Phonology (and its modern heirs) (Scheer 2011: §§139ff). The opening sentences appear under (38).

Kiparsky thus means to measure the distance between items recorded in the lexicon (morphophonemic representations) and items occurring on the surface (phonemic, i.e. after the application of all rules, or phonetic, i.e. in pronunciation), and then to assess what this distance may, or ought to be. The larger the distance, the more abstract is phonology; the smaller it is, the more concrete the phonology.

It is obvious that there is nothing to be measured if an alpha in the lexical recording were compared to a beta (produced by rule application based on an alpha) after the last phonological rule has applied (phonemic surface).¹⁵ In the same way, nothing could be measured at all when an alpha were compared to a [p], labiality or whatever other phonetic item. An alpha is not more or less distant from a beta, or from a [p]: it does not make sense to talk about distance between these items because they do not belong to the same type of vocabulary.

Kiparsky’s endeavour thus requires that items recorded in the lexicon and items appearing on the (phonemic or phonetic) surface both be written in a vocabulary that allows us to calculate their distance. This vocabulary is phonetic in kind. Kiparsky (1973 [1968]: 6) says that surface [ŋ] may be /ng/ lexically: in his view this is abstract, but not arbitrary, as would be a lexical origin /ö/ for [ŋ]. He argues that the two pre-generative approaches to phonology that he has identified either do not allow for abstract relationships (the Prague school and American structuralism, which he calls fully concrete), or allow for arbitrary relationships, which is the case of the Hjelmslevian view, “defended by Lamb (1966b), Fudge (1967) and, apparently, Householder (1965)” (p. 5), which he calls fully abstract (“[t]he writings of Hjelmslev may be interpreted as setting forth a similar position” p. 5). According to Kiparsky, then, generative phonology is a synthesis of the best of two worlds.

The presence of phonetic content in phonology is thus a prerequisite for Kiparsky’s project to measure abstractness. It is also the decisive property that distinguishes the new generative approach from the preceding fully abstract view represented by Hjelmslev. In this context, it is of note that Kiparsky explicitly mentions Jakobson (1948) (where the author admits to be a sinner against de Saussure) as a forerunner of generative phonology.