The glottal stop is a bit of a misnomer, since it is usually not a stop. While other stops are also regularly lenited into fricatives or flaps, especially in prosodically weak positions (Mitterer & Ernestus, 2006; Warner & Tucker, 2011), the glottal stop is special in that not achieving a full closure is the default. Ladefoged and Maddieson (1996, p. 75) write: “In the great majority of languages we heard, glottal stops are apt to fall short of complete closure, especially in intervocalic positions.” Interestingly, the only other somewhat similar phoneme is the glottal fricative [h], which is often produced as a breathy vowel rather than a fricative (Garellek et al., 2023). Indeed, we will see how the glottal stop and the glottal fricative are often partners in crime that differ from phonemes with primarily oral constrictions.

The glottal stop, like most other speech sounds, is a phoneme in some, but not all, of the world’s languages. A search of the UPSID database (Maddieson, 1984) shows that about half of the languages contain the glottal stop as a phoneme. That is not unusual. The voiced velar stop /g/ is in a similar ballpark, occurring in about 60% of the languages in the UPSID database. What makes the glottal stop interesting is what happens when it is not a phoneme, as, for example, in English. Here it can be used to mark prosodic boundaries and/or prominence on vowel-initial words (Dilley et al., 1996; Garellek, 2013; Steffman, 2023), signal stop voicing for stops in coda position (Penney et al., 2018), and replace oral stops (Schleef, 2013). This is unusual; I am not aware of languages that do not use [g] as a phoneme but then use [g] to mark prosodic boundaries or replace other stops (unless through voicing or place assimilation). Similarly, the glottal stop is sometimes considered to be the source of the /h/-aspiré in French (Boersma, 2007), in which phonetically vowel-initial words trigger phonological processes as if they were consonant-initial. The glottal stop also appears as the stød in Danish, where it replaces the pitch accents that occur in cognates from the related languages Norwegian and Swedish (Grønnum, 2023).

Moreover, the glottal stop is also an interesting case because its status in a given language is often ambiguous. In German, for instance, the majority opinion seems to be that the glottal stop is not a phoneme (Wiese, 1996), though other opinions can be found (Maas, 2006). Wiese (1996, p. 58–61) argues that it is not necessary to grant the glottal stop phonemic status, since its presence can be inferred by a rule that prefixes any vowel-initial foot with a glottal stop. It is important to note that this argument is purely grammatical in the sense that it focuses on what is needed to correctly produce German words.

The data presented here suggest that the glottal stop may very well be a phoneme in German, especially given that the glottal stop does not seem to have a lot of phoneme strength in Maltese, a language where it is clearly a phoneme. As discussed below, the Maltese glottal stop even defies the typical examples of a weak phoneme in the typology for marginal contrasts suggested by Hall (2013).

The relation between orthographic representations (in an alphabetic script) and phonological representations is one that is mostly discussed in the field of visual-word recognition and reading acquisition (Morais et al., 1979; Ziegler, 2018). For most adult readers, it comes as a surprise that the appreciation of speech as a concatenation of segments is not something that is obvious to everyone. In fact, Morais et al. (1979) had to show that empirically (see also Morais, 2021). They identified two groups of adults that had not learned to read as children. One group, however, had learned to read in adulthood while the other remained illiterate. This type of sampling helped to make the two groups comparable in terms of socio-economic status. These two groups were then asked to perform phoneme-level manipulation tasks such as phoneme deletion (e.g., bread minus ‘b’ = red). The results were so clear that the paper could be published without a statistical test of this result; the adults with reading training vastly outperformed the illiterate participants.

One upshot of this study was the question of whether reading training only reflects meta-linguistic performance or whether it is, in fact, necessary for phone-sized representations to arise in spoken-word recognition. This issue touches upon a long-standing debate in the field of spoken-word recognition; namely, to what extent the acoustic input is transformed before lexical access is attempted. And since a key must fit its lock, assumptions about transformations at the pre-lexical level have repercussions for the form of lexical entries in the mental lexicon. For instance, in exemplar-based models of speech perception and phonology (Pierrehumbert, 2002), if the pre-lexical level only generates a grainy spectrogram—that is, only a minimal transformation of the input—the lexical entries need to be stored in that format to match input to stored representations. On the other end, models of spoken-word recognition like TRACE (McClelland & Elman, 1986) or Shortlist (Norris, 1994; Norris & McQueen, 2008) assume that lexical representations are strings of phonemes, akin to classical linguistic theories (cf. Kucera, 1983). If that is the case, the pre-lexical system needs to generate a string of phonemes from the acoustic input so that lexical access can be achieved.

At this juncture, the findings of Morais et al. (1979) could be interpreted as follows: Pre-readers use a holistic approach for word recognition, as in Klatt’s (1989) suggestion of Lexical access from Spectra or the previously mentioned exemplar-based models of spoken-word recognition. By learning to read, however, the spoken-word recognition system is re-organized and now uses segments, as in TRACE (McClelland & Elman, 1986) or Shortlist (Norris, 1994; Norris & McQueen, 2008). This most radical interpretation of these results has been refuted by the finding that pre-readers (at a pre-school age) can learn about segment-sized units in speech perception (McQueen et al., 2012). Moreover, learning to read does not seem to modify this ability strongly, because McQueen et al. found similar patterns for 12-year-olds as for pre-readers.

However, one interpretation that remains is that orthographic units are necessarily activated in spoken-word recognition. This idea is supported by the finding that inconsistencies between orthographic and phonological form lead to a slowdown in the recognition of spoken words (Rastle et al., 2011; Ziegler & Ferrand, 1998). In this context, scholars from the field of visual-word recognition have focused on inconsistencies between orthographic and phonological forms that arise on the side of orthography (in words such as yacht, knight, pint, etc.), overlooking inconsistencies that occur frequently in the speech signal, at least in normal conversation. Here, every other word contains a phonetic change that makes the word inconsistent with its canonical form (Johnson, 2004), which is the one usually represented in orthography (e.g., plice for police, prowly for probably, yeshay for yesterday). This would mean that the requirement of a match between phonological input and orthographic representations would come at a massive cost for spoken-word recognition in normal conversation, casting doubt on the usefulness of such a requirement.

From the side of spoken-word recognition, it has been argued that the match with the orthographic form may boost the recognition of full forms that are relatively infrequent, the so-called canonical-form advantage. This effect indicates, for instance, that words with medial /t/ in American English produced with a full /t/ are recognized faster than the variant with a flap (Connine et al., 2008; Racine et al., 2014; Ranbom & Connine, 2007), even though the latter form is more frequent. Note, however, that influences of orthography on speech perception need not stem from an online activation of orthographic forms. It has been shown that reading a novel word leads to a phonological representation (Bakker et al., 2014) This indicates by extension, reading a known word will strengthen its phonological representation and in particular, its canonical-form variant. This, in turn, can indirectly influence spoken-word recognition without the need to be activated during spoken-word recognition.

While it has been found many times that reduced words are recognized slower than their canonical counterparts (Brouwer et al., 2013; Ernestus et al., 2002; Pitt et al., 2011; Ranbom & Connine, 2007), this does not necessarily mean that orthography is involved. Clearly, the canonical form provides the best evidence for the intended word. An analogy with object recognition might be helpful here; the canonical side view of a road bike provides the best information for recognizing a road bike (Edelman & Bülthoff, 1992). One would hardly be surprised to learn that visual objects are less well recognized if critical parts are occluded or not well visible from the current angle. Similarly, a reduced word form with missing segments is less efficiently recognized than its canonical counterpart. This canonical-form advantage (for a discussion of whether the canonical-form advantage is real, see Bürki et al., 2018) hence cannot be decisive in this debate.

At this juncture, the glottal stop provides an interesting window on this question. Since the glottal stop in German is part of the canonical form but not orthographically coded,1 its deletion should hinder spoken-word recognition to a lesser degree than the reduction of a comparable segment that is orthographically coded, if orthography is indeed automatically activated during spoken-word recognition. As comparable segments, Mitterer and Reinisch (2015) chose the Maltese glottal stop as a segment represented by the grapheme <q> and the German /h/. German /h/ is comparable to the glottal stop with regard to its restriction of occurrence (only pre-vocalically) and the only other segment that—in this position—is, with some regularity, deleted in conversational speech. An analysis of the Kiel Corpus showed that the likelihood of deletion is similar (see Mitterer & Reinisch, 2015).

To measure the effects of deletion of these segments (German /Ɂ/, Maltese /Ɂ/, and German /h/) on lexical access, Mitterer and Reinisch (2015) used a visual-world paradigm. Participants heard sentences produced in a casual manner (e.g., That guy simply ignored the red light) in which target words (light = Ampel /(ʔ)ampəl/ in German) were produced with or without the initial segment. To make sure that the targets only differed in the absence or presence of these initial segments, each sentence was recorded multiple times, sometimes with the instructions to delete the initial segment. Since this instruction may lead to generally more reduced productions of the target word, the cross-splicing made sure that the production of the non-initial segments, as well as the sentence context, was comparable across conditions. Note that acoustic confounds may be an important factor for other findings, as well. When materials for a lexical decision task are prepared, it may be the case that difficult-to-read words (such as yacht) are produced somewhat less clearly than consistent words. In our case, one base recording was used with the critical word juncture spliced in from a deleted or realized recording (see Table 1), so that the body of the target word was the same across conditions.

The results were clear. The deletion of all three segments inhibited lexical access; participants took longer to fixate on the target word (represented by an image, so that orthography was irrelevant in the task) if the initial segment was deleted. These reduction costs were equivalent for all three cases. Note that the experiment contained two tests of the critical hypothesis that the deletion of an orthographically coded segment is more harmful than that of an uncoded one; but both tests came to the same conclusion.

At this stage, one may wonder whether the test results are due to some idiosyncrasy in Mitterer and Reinisch’s (2015) materials. To test this, another experiment was run with a meta-linguistic, explicit task and clear-speech materials. Participants were asked to rate how well produced were the target words carrying German /Ɂ/, Maltese /Ɂ/, or German /h/ when presented in a short phrase. The same cross-splicing technique (see Table 1) was used to make sure the conditions only differed in the absence or presence of the initial segment. Unsurprisingly, ratings were lower for the deleted versions compared to the full versions. However, this time there was an orthographic effect so that the deletion costs were smaller for the uncoded German glottal stop than for the coded segments. This indicates that orthography does not influence implicit phonological processing during lexical access but does influence explicit processing.

Hence, perhaps orthography does not influence phonological processing, but it may have influenced phonologists who granted phonological status to German /h/ but not German /Ɂ/. As the forward slashes in the previous sentence indicate, I take these results as indicating that the glottal stop in German should be considered phonemic, especially given that /h/ has phonemic status. Appealing to the efficiency of storage, one common argument against glottal stop as a phoneme is that its occurrence can be explained by a post-lexical process (Wiese, 1996). As it turns out, this computability argument against the glottal stop as a phoneme is problematic, since the same argument could be applied to /h/ instead. Consider the following: Once we assume that all vowel-initial feet are glottal-stop initial, we can formulate a rule that /h/ is superfluous as a phoneme. We can then assume that [h] is generated by a rule that supplies vowel-initial feet with a glottal fricative. Now the choice between /h/ and glottal stop as a phoneme in German has become arbitrary, and the empirical data (deletion likelihood, importance for lexical access) do not provide evidence for one over the other. Moreover, we can assume a glottal stop phoneme also takes away the awkward possibility that /h/ and /ŋ/ are allophones of the same underlying phoneme that is realized differently in onset and coda, since /h/ would no longer be the only candidate allophone of /ŋ/ restricted to the onset.

Because of this, lexical representations of German words which are typically considered underlyingly vowel-initial have to be considered glottal-stop initial. Moreover, even though deletion of the initial segments delayed looks to target words, they were still quickly recognized; that is, the effect size of deletion on recognition was small. This is best explained by assuming multiple lexical form representations for the same word, so that the typical German mental lexicon would contain forms such as [ʔɑimɐ] and [ɑimɐ] for recognizing both variants of the German word Eimer (bucket). This gives rise to the question of how to conceptualize reduction and deletion of the glottal stop in production. Bürki and Gaskell (2012) provide experimental tools to investigate this issue in production. They investigated schwa deletion in English and asked participants to name pseudo-homophones—nonword letter strings that sound like words—such as kamra for the word camera. The underlying auxiliary assumption is that naming latencies will be shorter if the pseudo-homophone activates the lexical entry. This latency advantage was observed for schwa-bearing and schwa-absent pseudo-homophones of words with post-stress schwa (such as camera), but only for schwa-bearing versions of words with pre-stress schwa (such as salami). Bürki and Gaskell argued that their results suggest that sometimes there is only one representation and deletion occurs during production (for pre-stress schwa words like salami), while for other words (post-stress schwa words such as mackerel), speakers chose between two possible versions. These results lead to the open question of which model better applies to glottal-stop initial (and /h/-initial) words in German.

Independent of the phonological question of phonemic status, the data clearly show that, for the purpose of language processing, /h/ and glottal stop are of similar importance for lexical access in German. This gives rise to the question of how L2 learners, especially those with a Romance L1, like Italian,2 learn this pair of confusable consonants. To our knowledge, little research about this situation exists; most of it focuses on cases in which one member of the L2 pair is similar to an L1 phone (Bradlow et al., 1997; Cutler et al., 2006; Ju & Luce, 2004; Weber & Cutler, 2004).

Theories of L2 acquisition focus on the similarity of L2 phonemes to L1 phonemes (Best & Tyler, 2007; Flege & Bohn, 2021), and, for contrast, explore the question of how they map onto the L1 system. The two glottal consonants of German are the relatively rare case in which both sounds would be, in terms of the Perceptual Assimilation Model (Best et al., 2001), non-assimilable sounds for learners with a Romance language background. French, Italian, and Spanish L1 learners miss both glottal consonants in their L1 inventory and there are no obvious assimilation candidates in their L1 inventories. Therefore, we (Eger et al., 2019) decided to investigate how Italian listeners deal with /h/ and glottal stop in German. While there are many studies looking into the acquisition of /h/ by Romance learners (e.g., White et al., 2017), most of these study /h/ exclusively and do not consider the glottal stop. As our data show, this leads to an incomplete picture.

From a psychological perspective, investigating the glottal stop in L2 acquisition of German lends insights into the implicit learning abilities in language acquisition after puberty, because the glottal stop is not coded and most formal L2 acquisition curricula do not go beyond the phonological properties of the written language (which excludes the glottal stop). This, therefore, provides a window on the implicit learning abilities in language acquisition after puberty. Note that some scholars argue that implicit learning is severely impaired, if not impossible, after puberty (Clahsen & Felser, 2006). Moreover, the comparison of glottal stop and /h/ contributes to the discussion of (undeniable) orthographic influences on L2 phonological processing (Bassetti, 2017; Hayes-Harb et al., 2018; for a review, see Hayes-Harb & Barrios, 2021).

Following Mitterer and Reinisch (2015), Eger et al. (2019) tested how glottal consonants affect L2 speech processing in an explicit rating task and the visual-world paradigm. However, instead of presenting the target words with the glottal consonant present or deleted (as in Mitterer & Reinisch, 2015), Eger et al. presented substituted tokens. That is, taking the example of im Eimer (‘in the bucket’) presented in Table 1, they presented the correct pronunciation [imʔɑimɐ] and a substituted version in which the glottal stop was replaced by /h/, [imhɑimɐ]. In addition, Eger et al. also used a production task. To elicit /h/- and glottal stop-initial words, but without presenting orthographic cues, participants were guided to produce sentences based on a visual representation. Participants were familiarized with some actors (taken from the animated series Peanuts), images representing common actions (a set of coins for buying), a number (e.g., 9) and a visual representation of an /h/- or glottal stop-initial word (e.g., the picture of a hat), with the instructions to generate a sentence from these cues. Given the cues ‘Lisa’, ‘9’, ‘hat’ and ‘buy’, the resulting sentence would be “Lisa has bought nine hats” (Lisa hat neun Hüte gekauft, [ˈliːza hat nɔʏn ˈhyːtə gəˈkaʊft]; note that in German, the main verb is sentence final if there is an auxiliary verb). The Italian participants were living in Germany and were intermediate to advanced (levels B1 or C1) speakers of German, and they had no problem performing this task.

Especially for glottal-stop initial items, we might expect that the Italian learners would simply produce these items as vowel-initial. Note that the sentences avoided a hiatus, since the words preceding the vowel-initial word ended on a consonant (e.g., neun, ‘nine’). The results, however, showed a relatively successful acquisition of both glottal stop and /h/ by Italian learners, with about 65% correct productions for glottal stops and 75% for /h/. This difference was not significant. Given the small sample size of 13 learners, this should not be overinterpreted; however, this comparison is within participants, so strong effects would likely materialize. And here we should expect a strong effect; after all, teaching materials, such as dictionaries (e.g., Macchi, 1985), instructed learners to produce /h/, but not glottal stop. Moreover, the orthography constantly reminds learners of /h/, while the glottal stop receives no advantage from an explicit orthographic representation. Despite these combined pressures favouring /h/ production, the observed difference between the two sounds remained relatively small.3 This result suggests that learners can readily acquire glottal-stop usage through implicit learning.

Among the incorrect productions, which consisted of deletions and substitutions (the word ‘Alder’, ‘eagle’, produced as Hadler), Italian learners more often deleted the glottal stop than /h/, while substitutions were equally likely for both targets. Again, these results indicate surprisingly little difference between glottal stop and /h/, given the massive difference between these segments in terms of teaching and orthography.

Based on this relatively strong performance in the production task, the results of the visual world eye-tracking task were surprising. Italian learners were not influenced at all by substitutions. They fixated on the target images regardless of whether words were presented in their correct or substituted pronunciation. That is, fixations on the picture of an eagle did not depend on whether the word was produced correctly as [Ɂadlɐ] or incorrectly as [hadlɐ]. Unsurprisingly, this contrasted with the results of the German control group, who showed strongly delayed fixation to the target picture when the target had the substituted consonant. However, in both datasets, the effects were symmetric; the substitution of a glottal stop with [h] yielded similar effects to the substitution of [h] with a glottal stop. These results are indeed difficult to reconcile with the assumption that /h/ has phonemic status, and the glottal stop is an epenthetic segment. They hence support the earlier claim that the two segments have a similar status in German.

The results obtained with Italian learners of German show a strong dissociation between perception and production. Usually, one expects that perception must lead production, since learners first need to perceive a difference before they can produce it. But our findings show the opposite pattern. However, this overlooks the issue that perception may not be a unitary phenomenon. Neuro-imaging data (Scott & Wise, 2004) indicate that pre-lexical perception may consist of two streams, one for inferring the speech gestures that produced the sound (a how? pathway) and a stream for word recognition (a what? pathway). The data can be explained by assuming that the how pathway has, at least partly, mastered the distinction, although this distinction is not used in the what pathway. In line with this interpretation, the results of the rating task indicated that, when asked explicitly, potentially engaging the how pathway, Italian learners distinguished between correct and substituted utterances, that is, they rated [Ɂadlɐ] as a better pronunciation of Adler than the incorrect [hadlɐ]. That is, they can perceive the difference explicitly, and this ability can then be used to guide production.

However, for word recognition, the glottal consonants are not used to distinguish between words for Italian learners of German. This can be explained by assuming a) Italian learners represent glottal-stop and /h/-initial words as vowel-initial (i.e., /adlər/), or b) they have one category for glottal consonants in their perceptual lexical representations. If the latter is the case, words that are orthographically vowel- or /h/-initial are stored in the mental lexicon as containing some glottal marking that is not specified to be a stop or a fricative (i.e., /(^h/_ʔ)adlər/). An additional visual-world eye-tracking experiment was run to distinguish between these two accounts. In this experiment, /h/- and glottal stop-initial words were presented in the auditory stream as (phonetically) vowel-initial words. If Italian learners do represent these words as vowel-initial, without any glottal marking, they should be just as unaffected by the deletion of /h/ and glottal stop as they were by their mutual substitution in the first eye-tracking experiment. However, the results revealed a difference between correct and deleted productions, which shows that Italian learners represent these words as containing an initial glottal consonant. However, this consonant can either be a glottal stop or a glottal fricative.

Overall, the results from this project indicate that in L2 learning, implicit processing is surprisingly effective. For instance, in perception, learners learn that German does not have (phonetically) vowel-initial words, and they learn the same for glottal-stop initial words, even though the glottal stop is neither represented orthographically nor taught explicitly in L2 pedagogy. (Eger et al. interviewed an academic who trains Italian aspiring teachers of German to verify this.) Similarly, in production, there are only small differences despite the strong differences in how well /h/ and glottal stop are emphasized during L2 acquisition.

Mitterer (2018) investigated the cues to gemination in Maltese, making use of the minimal pairs that arise in the verb paradigm of Maltese. As a Semitic language, Maltese uses trilateral roots (such as d-ħ-k, ‘laugh’), and these roots take on different meanings in different forms. The verb form that is useful for this case is the second verb form, which gives a causative meaning. That is, John daħak means ‘John laughed’, but Agnes daħħak John means ‘Agnes made John laugh’. To elicit such forms without reading, the study made use of a sentence guessing game, in which participants saw one or two actors and, optionally, a third participant or object plus the root in written form (e.g., r-q-d). They then had to guess the sentence (e.g., for a root with a medial glottal stop such as r-q-d, ‘sleep’: John raqad fuq sufan, ‘John slept on a sofa’; Agnes raqqad lit-tifla, ‘Agnes brought the girl to bed’). The results showed, unsurprisingly, that the geminates were longer than the singletons. However, the effect size of this difference differed over segments, so that singleton versus geminate oral stops were better separable by duration (d = 1.7) than glottal stops (d = 1.1). However, as suggested by Ladefoged and Maddieson (1996), there was an additional cue for the glottal stop, so that more than 80% of the singleton glottal stops were, in fact, not full stops but periods of glottalization, while more than 80% of the geminate stops contained a full closure. An additional study tested whether listeners made use of this secondary cue in perception. This study used a simple Two-Alternative Forced Choice (2AFC) task with three duration continua. One continuum contained a glottalization that never achieved full closure, and this glottalization was made successively shorter by cutting out full periods of the vocal fold cycle. In a second continuum, half of the glottal cycles were set to zero, mimicking the on-off voicing pattern that is often observed with glottal stops, and in the third, there was a full closure for the whole duration of the glottal stop. While listeners still used duration as a cue to distinguish singleton and geminate glottal stops, they were highly sensitive to these secondary cues. In fact, listeners found it difficult to accept a long glottalized period as an example of a geminate glottal stop, despite a duration that was typical for geminate consonants.

Similar results were obtained for /h/, which tends to be produced with an additional oral constriction when used as a geminate. Importantly, no such secondary cues could be found for alveolar stops and fricatives. Although the center of gravity of /s/ was slightly higher in geminates, the perception study found that listeners ignored this cue and focused on duration alone. These findings are difficult to reconcile with accounts that assume that gemination is perceived as a unitary feature, possibly prosodic in nature (Kotzor et al., 2017), because the cues at the pre-lexical level are too diverse to be recognized as a single feature. The finding that gemination may therefore not be recognizable from the input as a feature dovetails well with another line of research that questions the use of features in perception, namely, that learning for speech perception (Norris et al., 2003) fails to generalize in ways predicted by feature-based theories (Mitterer et al., 2016; Mitterer & Reinisch, 2017; Schumann, 2014).

As already mentioned above, Maltese uses the glottal stop not only as a phoneme—which will be called “lexical” glottal stops—but also at the onset of otherwise vowel-initial words. If vowel-initial words are marked with a glottal stop, this could (from a theoretical point of view) lead to disastrous consequences. Suppose vowel-initial words are stored with a single lexical representation, in line with the canonical form without a glottal stop. For example, attur /at:ur/ (‘actor’) would be stored without an initial [ʔ]. Now the input [ʔa] activates a whole cohort of /ʔa/-initial words (qattus, qabad, qabel, qabeż, etc.) and deactivates the vowel-initial word. Bottom-up mismatch has strong consequences in spoken-word recognition: In order to model lexical-decision data in the original Shortlist model (Norris, 1994), the penalty for a mismatch was as strong as the activation caused by three matching segments. For example, consider the activation of the word lighter /laɪɾɚ/ upon hearing writer /ɹaɪɾɚ/. The initial mismatch would lead to an activation level of –3. That is, the lexical-decision data, on which Shortlist is partly based, indicate a strong penalty for mismatch. Additionally, the cohort of /ʔa/-initial words would, through lateral inhibition, also deactivate the target attur. This would mean that such words would only be recognized relatively late, when all other lexical competitors starting with /ʔ/ would be deactivated by mismatch with the input. Clearly, this is unlikely to happen. Maltese speakers do understand Maltese, and the use of epenthetic glottal stops is ubiquitous.

There are two main ways in which this problem can be solved. As in many areas of cognitive science, the key question is whether the problem is solved by representation or by computation. A representational solution is straightforward: The mental lexicon has more than one form representation for vowel-initial words, so that for a vowel-initial word (e.g., attur), the mental lexicon contains the forms /ɑt:ur/ and /ʔɑt:ur/. For the current purpose, it does not matter whether these are fully episodic, grainy spectrograms (Goldinger, 1998; Klatt, 1989) or more abstract representations based on segments (Bürki & Gaskell, 2012). In this case, recognition is straightforward. The input [Ɂɑt:] activates both the words qattus and attur based on the full overlap with (one of) their lexical representations. When the final segment [r] arrives, qattus is deactivated and a full match emerges with /ʔɑt:ur/. The processing solution, on the other hand, assumes that listeners might distinguish an epenthetic from a lexical glottal stop. A similar claim has been made for the difference between a place-assimilated [p] in cat brought and an intended [p] in cap (Gow, 2003).

Mitterer et al. (2019) investigated this issue in a study with both production and perception tasks. During the production task, participants produced sentences from a prompt that backgrounded the target words. This was done to ensure participants produced the target word and its preceding context fluently, without pausing before the critical word. Pausing makes the detection of a glottal stop difficult. Figure 1 shows an example of a production prompt. Participants were instructed to answer the question shown in red (‘Did Anna say the word art in this case?’), based on the display. The speaker, however, is Nina. Participants had been familiarized with four speakers and were reminded of their names by the letter on the t-shirt. The correct answer in this case is Le, NINA tgħid il-kelma art f’dan il-kaz (No, NINA said the word art in this case’). The capitalization indicates contrast on who is speaking, while the target word is deaccented due to its given status. In this case, the target word is preceded by a vowel, but in other cases, the vowel-initial target word could be preceded by a word ending on a consonant. This design makes it possible to test whether the epenthetic glottal stop is used to prevent vowel hiatus, as Booij (1995) assumed for Dutch.

The results were surprising in every respect. Despite the prosodic backgrounding of the vowel-initial words, half of them were marked with an epenthetic glottal stop, suggesting that epenthetic glottal stops are indeed ubiquitous in Maltese. Secondly, an epenthetic glottal stop on the vowel-initial word was not more likely in case of a hiatus (kelma, art) than in case of a preceding consonant (kliem, art). None of the measured phonetic properties—such as duration or degree of closure—showed a difference between epenthetic and lexical glottal stops. However, an exploratory analysis found that the duration of the preceding word was longer when vowel-initial words were marked with an epenthetic glottal stop. This was interpreted as the effect of domain-final lengthening. This indicates that the Maltese epenthetic glottal stop is—just like in English (Dilley et al., 1996)—used to mark larger prosodic boundaries. However, it is important to note that this difference was relatively small. Given the small effect size, it is unlikely that listeners can clearly distinguish between epenthetic and lexical glottal stops based on this cue alone. If listeners cannot reliably determine whether a glottal stop is lexical or not based on input properties, it follows that listeners need a lexical representation of vowel-initial words that includes the glottal stop. That is, if the processing solution does not work, only the representation solution remains.

The next logical step is to investigate whether listeners nevertheless use the one cue found in the production data—lengthening of the preceding word—to interpret whether a glottal stop is likely to be lexical or epenthetic. Other research had already shown that listeners take prosodic properties into account when judging segmental information (Kim & Cho, 2013; Steffman, 2023). We therefore tested this possibility via a 2AFC task in which participants had to decide whether a carrier phrase contained the word għam (/ɑ:m/) or qam (/Ɂɑ:m/). The stimuli differed in the amount of phonetic evidence for the presence of a glottal stop and the duration of the preceding word—that is, its final lengthening. Participants took this latter prosodic information into account and interpreted the word more often as being underlyingly vowel-initial (despite the presence of glottalization) when the preceding word was produced with final lengthening.

The next question is whether this inference is fast enough to influence online word recognition. This was investigated with an eye-tracking study using printed words as targets (McQueen & Viebahn, 2007; Mitterer & McQueen, 2009). The use of written words as targets in the visual-world paradigm was necessary to facilitate the use of pseudo-onset overlap pairs such as qattus-attur (‘cat’-‘actor’) or qammiella-ammeta (‘a stingy person’-‘to acknowledge’), which share their initial segments if the vowel-initial word is produced with an epenthetic glottal stop. Using only picturable items would have meant an unacceptably low number of usable items. The experiment made use of a straightforward two-by-two design with the factors of pre-boundary lengthening (the amount of final lengthening of word preceding the vowel- or glottal stop-initial target) and target identity (vowel-initial vs. glottal stop-initial). All target words were presented with a glottal stop, and cross-splicing was used, ensuring that the glottal stop realization for both members was identical. The results were somewhat disappointing, as none of the predictors had any measurable impact on the fixation patterns. However, one interesting finding was that the general pattern of looks to target and competitor patterned similarly to real onset overlap pairs (such as beaker and beetle, as in Allopenna et al., 1998), with this effect being symmetrical. That is, when hearing the word qattus, Maltese listeners look to the same extent at the words qattus and attur. This suggests that the input qatt… strongly activates the word attur. This is strong evidence for the claim that the mental lexicon contains multiple phonological-form representations for words that occur in multiple forms (such as Maltese vowel-initial words that are often produced with an epenthetic glottal stop).

This lack of an effect for pre-boundary lengthening on fixation patterns in the eye-tracking task was somewhat surprising given the effect of pre-boundary lengthening in the 2AFC task. An additional experiment using the pseudo-onset overlap pairs was therefore run as a gating task. This provides another test of whether the effect reliably arises during later processing (even though not during initial lexical access). This turned out to be the case as the initial effect was replicated with a 2AFC task in a gating task. Participants heard [ʔɑm:] in a carrier phrase, with the second part of the word masked by noise. The participants’ task was to guess whether the word was ammetta or qammiela. To make the task more feasible (and less frustrating) for the participants, two gates were used. On filler trials, a later gate containing disambiguating phonetic material was presented (i.e., a small part of the second vowel in the pair ammetta – qammiela). On the critical trials, only the shared initial segments were presented. As in the eye-tracking study, the carrier phrase could contain pre-boundary lengthening before the critical word. Listeners used this cue and assumed that the glottal stop was more likely to be epenthetic (that is, they indicated that the intended word was ammetta and not qammiela) when the preceding word had final lengthening. These data then indicate that prosodic information may only be used during later stages in spoken-word recognition (but see Mitterer et al., 2024; Steffman, 2023, with other findings).

Another way listeners might determine whether a glottal stop is lexical or epenthetic is by adjusting to a speaker. It has been well documented that listeners adapt to individual speakers (Norris et al., 2003; Roettger & Franke, 2019; Samuel & Kraljic, 2009). Therefore, an additional study was conducted (Mitterer et al., 2021a), in which participants were first exposed to a speaker who did or did not use epenthetic glottal stops (as a between-participant manipulation). After this exposure, participants were subsequently presented with the pseudo-onset overlap pairs in the visual-world paradigm, in which glottal-stop initial words were produced with a full glottal stop and vowel-initial words without any glottal gesture. The experiment revealed that listeners apparently cannot adapt to the way a speaker uses epenthetic glottal stops (possibly because epenthetic glottal-stop production might be variable within, but not between, speakers). A surprising result, however, was that glottal stop-initial words still showed a pattern of onset-overlap competition with vowel-initial words that were produced without any glottal gesture. That is, when hearing attur, there was a rise in fixation to qattus that was similar to the rise for attur, despite the absence of any glottal gesture in the stimulus. This result suggests that glottal stops, even when functioning as phonemes, do not strongly constrain lexical access. To fully show that this is the case, an additional experiment was run using /t/-initial words and vowel-initial words that showed similar pseudo-onset overlap if the initial /t/ is disregarded (such as tabib-abjad, ‘doctor’-‘white’). This experiment asked whether tabib would be strongly activated when hearing the target abjad, just as qattus is activated when hearing attur. As expected, this was not the case; there was no rise in looks to tabib when abjad was heard, and the pattern was significantly different from that with pseudo-onset overlap pairs that involved a glottal stop (i.e., qattus-attur).

This result brings us full circle as the glottal stop appears to function less robustly as a phoneme than other stops. Note that /t/ was not chosen arbitrarily, because it is generally considered to be the least marked segment and, hence, is the next most likely candidate to not strongly constrain lexical access. However, it constrains lexical access more than a glottal stop. That means that even in a language in which the glottal stop is a phoneme, it is less so than other phonemes. This sheds an interesting light on the case of German, where the glottal stop is often not considered a phoneme due to its low functional load and predictability. However, if the glottal stop is generally not a strong phoneme, it is not surprising that, in German, it doesn’t seem to carry a large functional load. Since this seems an inherent property of the glottal stop, this would constitute another argument to consider the glottal stop in German a phoneme (to the extent that a glottal stop qualifies as a phoneme).

In the previous section, the potential cross-talk between prosodic and segmental processing was discussed. Since prosodic properties can influence segmental detail (Cho et al., 2017), the question arises whether listeners not only use prosodic information to judge segmental properties (Kim & Cho, 2013; Steffman, 2023) but also use segmental properties to make inferences about the prosodic structure of an utterance. In English, where the glottal stop is mostly used to signal prosodic boundaries, it is hardly surprising that listeners use its presence to infer a prosodic boundary. However, this situation is different in Maltese, where the glottal stop must be processed as segmental information (because it is a phoneme). This leads to the following question: Can listeners then still use a glottal stop to infer the presence of a prosodic boundary when the glottal stop is not lexical but epenthetic?

Mitterer et al. (2021b) tested this by using coordinated names such as Daniel or Gordon and Malcom. Consider how this phrase could be interpreted as the answer to ‘Who should do the dishes?’ Daniel or Gordon and Malcom can then be interpreted as either ‘Malcom plus either Daniel or Gordon’ or ‘(poor) Daniel alone or Gordon and Malcom together’. In Maltese, this phrase is Daniel jew Gordon u Malcom. Note that an epenthetic glottal stop can only occur before the Maltese equivalent of ‘and’ u, since the Maltese equivalent of ‘or’ jew is not vowel-initial.

To test this in an experiment, such phrases were recorded and then manipulated along a duration continuum. At one extreme, the first name’s second syllable was long (Danielll) and the second name’s second syllable was short (Gordon); at the other extreme, the first name’s second syllable was short (Daniel) and the second name’s second syllable was long (Gordonnn). As expected, participants interpreted the phrases as having a boundary after the first name when the first name’s second syllable was long and the second name’s second syllable was short, but as having a later boundary after the second name when this pattern was reversed (first name short, second name long). This indicates that the paradigm successfully elicits prosodic judgements based on durations reflecting final lengthening. Moreover, if the question is whether the glottal stop is used, it is useful to provide other guiding information as well. Otherwise, there is the experimental affordance to use the only information in the experimental design (presence or absence of a glottal stop, see Durgin et al., 2009, for discussion of this issue), something that could be considered an overly confirmatory research strategy (see Firestone & Scholl, 2016, for more discussion of confirmatory research strategies).

In addition to these duration manipulations, the stimuli differed in the presence of an epenthetic glottal stop on u (‘and’). Participants could have ignored this information and focused solely on duration. However, they used the presence of the glottal stop to infer a prosodic boundary. That is, given the phrase Daniel jew Gordon u Malcom, participants were more likely to make a late-closure decision when the u was produced with an epenthetic glottal stop. Since late-closure is often considered to be the default parse of a phrase (Frazier & Rayner, 1982), a second experiment showed that epenthetic glottal stops can also support parses that go against the late-closure heuristic in phrases such as Daniel u Gordon jew Malcom, where the u follows the first name, rather than the second one. These data show clearly that segmental information can be used to guide prosodic processing.

These findings show interactions of prosodic and segmental processing streams. This suggests that theories of prosodic and segmental processing can no longer be treated as independent. Understanding the cross-talk between these two processing streams is a major task for future research (Steffman, 2021; Steffman et al., 2022), particularly because models must avoid an infinite regress—where prosodic decisions cannot be made without segmental information and vice versa.