1.1. Tashlhiyt and vowelless words

Tashlhiyt (iso: [shi]) is an Amazigh (Berber; Afroasiatic) language spoken in southern Morocco with an estimated 5 million speakers. The phoneme inventory of Tashlhiyt is consonant-heavy: there are 34 consonants (/b, m, f, t, tˤ, d, dˤ, n, r, rˤ, s, sˤ, z, zˤ, l, lˤ, ʃ, ʃˤ, ʒ, ʒˤ, j, k, kʷ, g, gʷ, w, q, qʷ, χ, χʷ, ʁ, ʁʷ, ħ, ʕ, h/), which contrast as singleton and geminate in all word positions, and three vowels (/a, i, u/) (Ridouane 2014). Tashlhiyt permits highly complex syllable structures, which are more likely to be found in languages with higher consonant-to-vowel inventories (Maddieson 2013; Easterday 2019). Phonotactics in Tashlhiyt are extremely permissive: words containing consonant sequences that challenge cross-linguistic sonority tendencies are common, such as word-initial sequences containing plateauing kti¹ or falling rku sonority profiles (Lahrouchi 2010; 2018; Jebbour 1996; 1999; Boukous 1987; 2009). Another cross-linguistically rare pattern in Tashlhiyt is the presence of “vowelless” words containing only consonants and no lexical vowels, e.g., tftktstt ‘you sprained it’ (Ridouane 2008; Dell & Elmedlouai 2012; though other varieties of Amazigh do display schwa epenthesis).

A very robust cross-language phonological tendency is for words to contain vowels. This is consistent with phonotactic sequencing accounts that syllables containing rises in sonority from periphery to nucleus are preferred (Clements 1990; Zec 1995). However, it has also been shown that constraints on sonority are language-specific and vowels-as-nuclei is not an absolute; for instance, some languages allow consonants to be syllabic. Even within the languages that allow consonant nuclei, sonorant consonants are the most commonly permitted syllabic segment (Bell 1978), and in many languages syllabic consonants are restricted to some environments (e.g., in unstressed syllables in German and English, e.g., bottle, button). Allowance of consonant nuclei in stressed positions (e.g., monosyllabic words) is less common, but can be found in languages such as Slovak (syllabic liquids) and Yoruba (syllabic nasals). The most cross-linguistically rare phenomenon is that of syllabic obstruents.

Tashlhiyt allows any segment to occupy the syllable nucleus (Dell & Elmedlaoui 1985; Bensoukas 2001; Ridouane 2008; Dell & Elmedlaoui 2012; Ridouane 2014; Lahrouchi 2018). Moreover, vowelless words are frequent in Tashlhiyt; for instance, Ridouane (2008: 328) reports that in a collection of texts, 22% of the lexical items are vowelless words and around 8% of words contain voiceless obstruents only. Lahrouchi (2010) compiled a list of 222 bi- and tri-segmental Tashlhiyt verbs and categorized them into classes based on the sonority profile of C1–C2 (note that a focus on word-onset sequences is consistent with work showing that the coordination of C1 and C2 in vowelless words is more constrained than that for C2 and C3 (Pouplier & Beňuš 2011)). About 20% of the verbs in Lahrouchi’s (2010) database were vowelless words containing an initial obstruent-sonorant sequence (rising sonority, e.g., krz ‘plow’), 17% contained an initial cluster of obstruents (plateauing sonority; e.g., gzm ‘cut’), 15% contained an initial sonorant-obstruent sequence (falling sonority; e.g., rgl ‘knock’). 33% contained vowels (e.g. knu ‘lean’) and just 2% contained plateauing sonority with initial sonorants (SSO or SSS). Thus, within tri-segmental vowelless words in Tashlhiyt, rising (OSC; S = sonorant, O = obstruent, C = any consonant),² falling (SOC), and plateauing (OOC) sonority profiles are all frequent forms. (Note that the morphology of the language is sensitive to the sonority structure of consonant-only verbs. A case in point is imperfective internal vs. external gemination, as in the monosyllabic verb krz/kkrz ‘plow’ and disyllabic verb mgr/mggr ‘harvest’ (Dell & Elmedlaoui 1985; Jebbour 1996; Bensoukas 2001).)

There is also a large body of work investigating the articulatory and acoustic properties of vowelless words in Tashlhiyt. Ridouane (2008; Ridouane & Cooper-Leavitt 2019) uses phonetic and phonological evidence to demonstrate that consonant sequences in Tashlhiyt are heterosyllabic and do not form clusters. In a complex cluster language, like English, consonant sequences form temporally overlapping, coarticulated gestures; for instance, in English, the duration of a consonant compresses when onsets contain more than one segment (e.g., the /k/ in scab is shorter than that in cab (Marin & Pouplier 2010)). However, in Tashlhiyt, as the number of consonants within a word increases, the segment durations remain stable indicating non-overlapping coordination of segments (Hermes et al. 2011; Ridouane et al. 2014; Hermes et al. 2017). Tashlhiyt listeners have been shown to treat consonant sequences as heterosyllabic: while participating in a game where they heard a lexical item containing an initial consonant sequence and then had to repeat back only the “first part”, Tashlhiyt speakers responded overwhelmingly with simpleton onsets, indicating that they parse consonant sequences into multiple independent syllables (Ridouane et al. 2014).

Furthermore, Tashlhiyt has been shown to have wide timing of sequential consonants (Fougeron & Ridouane 2008). In fact, in many languages that allow words with highly complex consonant sequences, the gestures for sequential consonants are timed far apart from one another (Gafos 2002 for Moroccan Arabic; Pouplier & Beňuš 2011 for Slovak; Tilsen et al. 2012 for Hebrew). This non-coarticulating-consonants property also appears to be a common feature of languages that permit vowelless words (Pouplier & Beňuš 2011). This often results in an intrusive acoustic element between consonants, ranging in quality from a vocoid to a full schwa-like element (Hall 2006; Kirby 2014). In Tashlhiyt, the vocoid or schwa does not increase the length of the sequence when it occurs (Ridouane & Fougeron 2011), compared to epenthesis in other languages where the addition of a schwa increases word duration (Hall 2006; Davidson & Roon 2008). The presence of these variable vocoid elements in consonant sequences in Tashlhiyt is argued to reflect a transitional acoustic signal due to production of consonants that are coordinated to be timed far apart from one another, rather than an epenthetic vowel (Ridouane 2008; Dell & Elmedlaoui 2012; though, cf. Coleman 2001).

1.2. Perceptual explanations for the rarity of vowelless words

The existence, and prevalence, of vowelless words in Tashlhiyt presents an opportunity to address a fundamental puzzle in phonological theory: why are vowelless words so rare across languages of the world? In the current study, we focus on possible perceptual explanations. Some argue that cross-linguistic phonological tendencies can be explained by perceptual processing mechanisms since many common phonological patterns have perceptual motivations (Ohala 1993; Blevins 2004; Beddor 2009; Harrington et al. 2019). Perception biases could be at play in the cross-linguistic preference for CV structure: it is harder to identify a consonant when it appears in a cluster (e.g., in English, listeners recognize the “b” in band faster than the “b” in brand or bland (Cutler et al. 1987)); and a speech stream of CV sequences will contain maximally distinct acoustic modulations which are more salient, and thus recoverable, to a listener than a speech stream containing only consonants or only vowels (Ohala & Kawasaki-Fukumori 1997).

Our approach follows theoretical frameworks exploring how the mechanisms of human speech processing can be a source of understanding the stability of sound patterns and how they emerge and evolve over time (Ohala 1993; Blevins 2004; Harrington et al. 2019). By investigating perceptual patterns of synchronic speech variants using laboratory methods, phonological forms that are harder for listeners to recover can be identified. As outlined above, there is much work examining the gestural dynamics of vowelless words in Tashlhiyt (e.g., Ridouane et al. 2014). However, investigating the perception of vowelless words is comparatively under-researched. We ask whether examining the perception of vowelless words in Tashlhiyt can provide clues as to why they are so cross-linguistically rare, and also perhaps, the phonetic pre-conditions that permit them to be maintained in Tashlhiyt.

One of the auditory mechanisms that might make vowelless words susceptible to mistransmission is that consonantal cues can be obscured when surrounded by other consonants (Wright 1996; 2004). Perceptual cues vary based on the neighboring sounds. This means that the “same sound” can be more or less perceptually recoverable depending on the context in which it occurs. For instance, stops generate robust formant transitions on adjacent vowels, due to the large amount of movement from a closed to an open oral constriction. Formant transitions are an acoustic cue that listeners rely on to identify both the place and voicing of plosives (Liberman et al. 1954; Benkí 2001). A stop produced before another stop will result in fewer and less robust acoustic cues in the speech signal: A release burst between the first and second consonants would provide the only cues for listeners about the identity of the initial segment; reduction, or loss, of the release burst could jeopardize listeners’ ability to recover the speaker’s intended form (Wright 1996). Thus, a potential perceptual bias against vowelless words comes from the role of acoustic cues in the maintenance of phonological contrast. In particular, we focus on the perception of Tashlhiyt tri-segmental vowelless minimal pairs contrasting in the middle segment.

Moreover, the perceptibility of vowelless words might vary based on sonority patterns. One view is that greater acoustic contrast between adjacent segments allows for more robust transmission from speakers to hearers (Ohala & Kawasaki-Fukumori 1997). If adjacent segments are acoustically similar, they are more likely to be confused by a listener, thus diachronically unstable. Ohala and Kawasaki-Fukumori predict that sound sequences containing plateauing sonority patterns (e.g., /sfs/) are the hardest to perceive since they contain the smallest differences in acoustic modulations across segments. Meanwhile, sequences containing larger acoustic perturbations (like, /ble/ or /ske/) are less susceptible to perceptual confusion, thus more phonologically stable. Evidence supporting this comes from Chen et al. (2022), who investigated Mandarin listeners’ discrimination of rising (e.g., kl), falling (e.g., lk), and plateauing (e.g., kp) onset clusters with and without vowel epenthesis. Discrimination performance was higher for both rising and falling sonority clusters, presumably because the acoustic difference between the consonants made them more perceptible to listeners, compared to plateauing clusters. For the perception of vowelless words, an acoustic similarity hypothesis is that since sequences of consonants are more acoustically similar, vowelless words will be overall harder to discriminate than words with vowels. Moreover, within vowelless words, the prediction is that forms with more similar sonority patterns should be even harder to discriminate. If these predictions are borne out, it could explain why vowelless words would not be robust to transmission from speaker-to-listener and, thus, evolutionarily dispreferred.

On the other hand, as outlined above, prior work on vowelless words in Tashlhiyt has found that they are often produced with vocoid elements between consonants (Ridouane & Fougeron 2011). Since Tashlhiyt exhibits low consonant-to-consonant coarticulation, the presence of robust cues in the speech signal means that the internal phonological structure of vowelless words could be recoverable by listeners. Some cross-linguistic evidence in support of this possibility comes from Wright’s (1996) study of stop+stop consonant clusters in Tsou across different word positions. He found that stop release bursts are absent intervocalically, but distinct word-initially where the formant cues are absent. He argued that speakers are actively enhancing secondary acoustic cues for consonants in contexts where they are more likely to be difficult for the listener to perceive; and indeed, when the release bursts are present in initial clusters, listeners accurately identify the stop+stop clusters. Thus, a cue preservation hypothesis is that the phonetic implementation of vowelless words in Tashlhiyt creates salient acoustic cues that make them perceptually stable.

It is also possible that perceptual patterns follow predictions made by a traditional sonority sequencing account (Clements 1990). One common way of assessing the perceptual preferences of consonant clusters is to present nonwords to participants (either auditorily, or through orthography) and have them rate the likelihood each item could be a possible word in their language. Berent et al. (2009, inter alia) have demonstrated that speakers from various language backgrounds prefer rising sonority consonant sequences over plateauing sequences, which are in turn preferred over falling sonority profiles (e.g., blif > bnif > bdif > lbif; rising > plateauing > falling). This is consistent with theoretical accounts that the sonority hierarchy governs sound sequences in an active way in synchronic grammars (Zec 1995).

1.3. Phonetic variation across speaking styles

Above and beyond their perception more generally, we also consider the impact on listeners when speakers explicitly enhance vowelless words. Talkers adapt their speech based on the communicative context: speakers vary between hyper- and hypo-speech variants depending on whether there is evidence that the listener will misunderstand them (Lindblom, 1990). Hyperarticulated speech contains a variety of acoustic enhancements relative to non-clarity-oriented speech, such as longer and more extreme segment realizations (Picheny et al. 1986; Krause & Braida 2004; Zellou & Scarborough 2019; Cohn et al. 2022). Also, the acoustic effects of clear speech have been shown to increase intelligibility for listeners (Picheny et al. 1985; Smiljanić & Bradlow 2011; Scarborough & Zellou 2013). While clear speech effects have been found across several languages (e.g., Smiljanić & Bradlow 2005 for Croatian; Kang & Guion 2008 for Korean; Tupper et al. 2021 for Mandarin), more work is needed to understand how enhancement affects different types of phonological contrasts.

Prior work on clear speech reports acoustic enhancements of both consonants and vowels (Krause & Braida 2004). Therefore, clear speech should improve the perceptibility of all words in Tashlhiyt. However, examining Tashlhiyt enhancement effects can further inform different hypotheses about perception of different vowelless word types. From an acoustic similarity account, words containing the largest acoustic differences across segments (i.e., words with vowel or sonorant centers) could receive the largest clear speech boost since phonetic enhancement might make those sounds even more acoustically distinct. Evidence for this comes from a recent study of the perception of onset singleton-geminate-cluster contrasts in Tashlhiyt (Zellou, Lahrouchi, & Bensoukas 2022). They found that discrimination of the rarer onset contrasts (geminate vs. non-rising onset consonant sequences) is harder for listeners in both clear and causal speech; in fact, there was no clear speech boost for contrasts involving the non-rising clusters. We predict similar patterns for vowelless words in the present study.

A cue enhancement prediction is that clear speech enhancements are targeted for phonological contrasts that might be particularly challenging for listeners. Wright’s (1996) observation of consistent release bursts in stop+stop clusters in contexts where other cues are not present in Tsou is consistent with this. There is also evidence that talkers selectively enhance phonetic cues in response to a listener misunderstanding a target word with a minimal pair competitor (e.g., Buz et al. 2016; Seyfarth et al. 2016). For the present study, if vowelless words containing plateau sonority are the hardest to perceive in reduced speech because the cues to segments are not salient, they could receive the largest perceptual boost in clear speech where the speaker is aiming to make words maximally intelligible to listeners.

Regardless of which hypothesis is supported, the mere existence of alternative forms of words due to hypo- and hyper-articulation provides a rich testing ground to examine the mechanisms of perceptually-based variation (Blevins 2004). Since vowelless words are typologically rare, we believe that a comprehensive investigation into their perceptual processing should also include systematic acoustic variants since listeners encounter multiple phonetic forms of words. Moreover, descriptively, examining how speech variation affects the range of different linguistic contrasts found across the world’s languages is important to understand the relationship between speech communication and phonology. Thus, we examine speech style variation as an additional factor in this present study.

1.4. Non-native speech perception

How do non-native listeners perceive vowelless words? There is much work examining how non-native listeners process unfamiliar consonant sequences. For instance, there is evidence that speech processing is sensitive to the sound sequencing probabilities in one’s native language (e.g., van der Lugt 2001; Best et al. 2007; Davidson 2011). Investigating cross-language speech perception of typologically rare contrasts adds to our scientific understanding about universal processing mechanisms of different types of speech sounds (Bohn 2017). For instance, cross-linguistically rarer phonotactic patterns are less likely to be rated as a possible word in one’s native language (Berent et al. 2009).

Comparing listeners with different language backgrounds can also inform about the fundamental mechanisms of speech processing. For instance, some researchers have argued that a near-universal in word identification is the possible-word constraint, which suggests that listeners exploit transitional probabilities when segmenting sequences of sounds into words. For instance, since most languages only have words that contain vowels, listeners from a range of different language backgrounds segment words more easily from a context where the remainder would itself create a word with a vowel (e.g., apple from vuffapple; vuff is an acceptable word form), compared to where the remainder would create a vowelless word (e.g., fapple; *f) (Norris et al. 1997). However, El Aissati et al. (2012) found that native Tarifiyt (an Amazigh language related to Tashlhiyt) listeners do not exhibit the possible-word constraint when segmenting words in running speech. They argue that fundamental speech processing mechanisms, such as the possible-word constraint, are simply not active during spoken word comprehension by native speakers of a language that allows vowelless words. We explore this possibility further by comparing native and non-native Tashlhiyt listeners’ perception of vowelless words in two additional types of tasks.

At the same time, there is work showing that clear speech increases the intelligibility of non-native sound contrasts (Kabak & Maniwa 2007; Zellou et al. 2022). Since L2 adult learners usually begin as naive listeners, exploring the perception of novel sound contrasts in a foreign language also has applications for second language acquisition (Bohn 2017). Looking at cross-language perception of vowelless words, as we do in the present study, is critical in order to make generalizations about what makes them harder (or not) to comprehend above and beyond what a native-speaker experience allows.

1.5. Current study

For the current study, we designed two experiments to investigate the perception of vowelless words in Tashlhiyt by native and non-native listeners. As mentioned above, vowelless words are frequent in Tashlhiyt and vary in length, from two segments (e.g., fk ‘give’) to multisyllabic words containing long strings of consonants (e.g., tsskʃftstt ‘you dried it (F)’, Ridouane 2014). In the current study, we focus only on tri-segmental vowelless words (CCC structure) as a test case. The present study had four experimental manipulations. We outline each of these manipulations in turn next, as well as how the perception of vowelless words can shed light on fundamental issues in phonological typology.

First, we varied the listeners’ language background: native Tashlhiyt speakers and speakers who are Tashlhiyt-naive (here, American English speakers). English does permit relatively complex syllable structures (e.g., strengths, sixths) and syllabic consonants (e.g., bottle). However, only sonorants in unstressed syllables can be nuclei (though, alternating with a schwa pronunciation, Roach et al. 1992), and no monosyllabic vowelless words are allowed.

Second, we used two different perceptual paradigms: a paired discrimination (41AX) task and a nonword acceptability ratings task. In a paired discrimination task (Fowler 1984), a listener hears two pairs of items - one contains the same word repeated twice, the other contains two different words. In our task, the different-pair items varied in the middle segment (e.g., tuf vs. tlf or ʁbr vs. ʁdr). The discrimination task gauges low-level, auditory processing of vowelless words since listeners rely on acoustic similarity between pairs. If the cross-linguistic rarity of vowelless words stems from an acoustic processing bias, we predict that discrimination will be lower for vowelless words than for words with vowels. Meanwhile, the word-likeness task requires listeners to compare the phonetic and phonological properties of stimuli to the characteristics of the words stored in their memory. It has been shown that nonwords with greater lexical support are rated as more word-like than nonwords with less (Munson 2001; Frisch et al. 2001). We predict that Tashlhiyt listeners will show greater perceptual sensitivity and higher word-likeness ratings of vowelless words than English speakers.

Thirdly, we examine variations in the sonority profile of the first and second segments in the tri-segmental items. We focus on this sonority profile since that is a meaningful classification for Tashlhiyt (Lahrouchi 2010). Also, word-onset sonority profiles are argued to be more constrained than coda profiles cross-linguistically (Pouplier & Beňuš 2011), and listeners are more sensitive to sequential probabilities in onset position (van der Lugt 2001). Tashlhiyt test words contained one of four sonority types (where V = vowel, C = any consonant, O = obstruent, S = sonorant): rising sonority with a vowel nucleus (CVC), vowelless with rising sonority (OSC), vowelless with plateau sonority (OOC; NB: SSC are not frequent forms in Tashlhiyt, Lahrouchi 2010), and vowelless with falling sonority (SOC). Each of these four sonority profile types are illustrated with examples in Figure 1. In each example, the segments are ranked based on the sonority hierarchy (Parker 2002) to illustrate the differences across word types. For rising sonority words (both CVC and OSC), sonority increases from the first to the second segment, crossing at least 2 ranks. In plateau sonority words, there is not a large sonority difference from C1 to C2. In falling sonority words, there is a large difference in sonority, but it decreases from C1 to C2. For the ratings task, we constructed Tashlhiyt-like tri-segmental nonwords varying in their structure. By investigating vowelless words that vary in their sonority profile, we can tease apart different theoretical accounts about their perception. On the one hand, as outlined above, it has been argued that the sonority hierarchy is an active mechanism in shaping the perception of sound sequences across languages (Berent et al. 2009). A prediction from such an account is that perceptual sensitivity within vowelless words decreases with decreasing sonority (rising > plateauing > falling). On the other hand, sonority difference has been argued to be more relevant for perception (Ohala & Kawasaki-Fukumori 1997). That prediction is that perceptual sensitivity will be higher when words contain greater sonority differences (e.g., rising & falling > plateauing). Moreover, differences across tasks can speak to low-level auditory processes (discrimination) vs. higher-level phonological mechanisms (ratings) in the perception of vowelless words.

Figure 1

Examples of Tashlhiyt words varying in their sonority profiles (V = vowel, G = glide, L = liquid, N = nasal, VDF = voiced fricative, VLF = voiceless fricative; VDS = voiced stop; VLS = voiceless stop; O = obstruent; S = sonorant; C = any consonant).

For our fourth manipulation, we elicited all the items for the perceptual experiments from a native Tashlhiyt speaker in both Clear and Casual speaking styles. The style manipulation allows us to explore the relationship between phonetic enhancement and phonological typology. Our predictions for how clear speech variation will influence the perception of vowelless words apply to the discrimination task most readily. One prediction is that vowelless words will receive less of a clear speech boost in perceptual sensitivity than words that contain vowels. Since vowels are the loudest, most sonorous segment, clear speech enhancement can make it even easier for listeners to discriminate words that have different vowels. On the other hand, if the phonetic implementation of vowelless words includes the articulation of salient cues to the internal phonological structure, we predict that the transitional cues that serve to carry information about the segment identities will be enhanced in clear speech. How might clear speech influence nonword acceptability ratings? To our knowledge, no prior study has examined the effect of speaking style on word-likeness judgements. The ratings task can reveal whether speaking style influences listeners’ acceptance of different types of nonwords.

Thus, the present study is a comprehensive investigation of the perception of words without vowels in Tashlhiyt. We examine lexical discrimination performance and nonword acceptability judgements of tri-segmental vowelless words containing different sonority profiles by native Tashlhiyt and English-speaking listeners. We also varied the speaking style of the words (clear vs. casual speech) to explore the effect of systematic phonetic variation on the perceptual patterns of Tashlhiyt words. We predict that the patterns across words, tasks, listeners, and speech style can inform theoretical understandings of the relationship between language-specific phonetic variation and phonotactic patterns, as well as how auditory processing mechanisms influence preferred word forms.

2.1. Methods

2.1.3. Acoustic assessment of the stimuli

Since prior work has shown that clear speech contains longer word durations (Krause & Braida 2004), we measured the durations of target words across speaking styles. Table 1 shows the mean and standard deviation of word durations for CVC and CCC items across styles. A two-tailed, unpaired t-test revealed that words are longer in Clear speech than in Casual speech (t(227) = 19.7, p < 0.001). We also asked whether vowelless words are overall shorter in duration since they do not contain vowels. A t-test indicated that CCC and CVC items are not different lengths (t(227) = 0.3, p = 0.76).

Table 1

Word durations in milliseconds (and standard deviations) for CVC and CCC items produced in clear and casual speech.

	Clear	Casual
CVC	437 (89)	245 (57)
CCC	428 (78)	257 (47)

Ridouane & Fougeron (2011) analyzed the acoustic properties of vowelless words in Tashlhiyt, focusing mainly on sequences of obstruents, and report that a vocoid-like element is often produced between consonants, ranging in quality from a release burst to a full epenthetic vowel. Figure 2 illustrates examples of the variation in the realization of vocoid-like elements across vowelless words produced for the current study. Vowelless words can be produced with no intervening vocoid element, as in the production of the word nʒħ ‘pass a test’ provided in Figure 2.A. In Figure 2.B, the production of the word ʒbd ‘pull’ is realized with a short vowel-like element (transitional vocoids are conventionally transcribed as [@] by Tashlhiyt scholars (Dell & Elmedlaoui 2012: 16) that has weak amplitude and formant structure. The production of the word zdm ‘collect wood’ in Figure 2.C is a full vowel, with longer duration than the element in 2.B. and clear, high frequency formant structure and large amplitude.

Figure 2

Waveform and spectrogram examples of vowelless words displaying no transitional element (nʒħ ‘pass a test’; Panel A), a weak transitional vocoid element (ʒbd ‘pull’; Panel B) and a transitional schwa vowel (zdm ‘collect wood’; Panel C).

We coded each vowelless word production for whether there was any inter-consonantal release or vocoid-like element either between C1 and C2 or between C2 and C3 (if there were two, we measured the duration of the longer element; this only occurred for one word: bdr ‘mention’). We additionally coded whether that element was a full schwa vowel, characterized by higher-level formant structure. We also measured the duration of any vocoid element as the temporal interval between consonants. Table 2 provides the percentage of vowelless words that contain a vocoid-like element, across speaking styles.³ The percentages of those elements that are full schwas across conditions, are also provided, as well as the duration of the element when present.

Table 2

Percentage of tokens with vocoid-like elements, percentage of tokens produced with a full schwa vowel (subset of vocoid-like element percentage), and mean durations in milliseconds (and standard deviations) of the element when present, in vowelless words across speaking style.

	% tokens with an inter-consonantal release or vocoid-like element	% tokens with full vowel (subset of vocoid presence proportion)	Intrusive element duration (when present)
Clear	60.7%	39.3%	.034 s (.05 s)
Casual	53.6%	26.8%	.019 s (.02 s)

We ran tests to examine whether the rates of either any vocoid or the full schwa differs across speaking styles. A logistic regression run on presence of vocoid (=1) or not (=0) revealed that the likelihood of a vowelless token being produced with a vocoid-like element is not different across clear and casual productions [est. = 0.15, z = 0.76, p = 0.4]. A separate logistic regression run on presence of full schwa (=1) or not (=0) also reveals that the proportion of vowelless words with full vowels is not different across clear and casual productions [est. = 0.29, z = 1.4, p = 0.2]. However, a linear regression demonstrated that duration of the vocoid-like element, when present, is likely to be longer in clear speech productions [est. = 0.01, t = 2.1, p < 0.05]. Thus, vocoid-like elements, or full epenthetic schwas, are often produced in vowelless words (consistent with Ridouane & Fougeron 2011; Ridouane & Cooper-Leavitt 2019). While vocoid elements occur at the same rate in clear and casual speech, they are more likely to be longer in clear speech when they are present.

2.2. Results

2.2.1. Overall patterns

The data from the 72 listener responses were coded for correctly (=1) or incorrectly (=0) selecting the pair that contained different words in each trial and modeled using a mixed effects logistic regression with the glmer() function in the lmer package (Bates et al. 2015) in R. The model included three fixed effects (all treatment coded): Language Background (Tashlhiyt, English [reference level]), Speech Style (Clear, Casual [reference level]), and Minimal Pair Type (Non-minimal pair [reference level], CVC vs. CVC, CVC vs. CCC, CCC vs. CCC). The random effects included random intercepts for each participant, word pair type, and word pair ordering, as well as by-participant random slopes for Speech Style (including by-participant random slopes for Minimal Pair Type resulted in a singularity error indicating over fitting).

The addition of 2- and 3-way interactions between predictors was evaluated via model comparison using the anova() function based on the Akaike Information Criterion (AIC), measuring model goodness of fit while penalizing over-parameterization. The model that results in the lowest AIC is best supported by the data (Akaike 1974). The model with two-way interactions between Minimal Pair Type and Style as well as Language Background and Style had the lowest AIC (AIC = 3732.1, model ANOVA Chisq = 7.4, df = 1, p < 0.01) and was thus retained. Inclusion of the three-way interaction did not improve model fit. To explore significant interactions, Tukey’s HSD pairwise comparisons were performed within the model, using the emmeans() function in the emmeans R package (Lenth et al. 2021).

Figure 3 shows mean discrimination performance for each Minimal Pair Type by Speaking Style, across listener groups. Even though there was not a 3-way interaction, Figure 3 displays mean discrimination accuracy for each Minimal Pair Type across Speech Styles for each Listener Group so that the parallel patterns of performance for Tashlhiyt and English listeners can be observed. The full model output is provided in Table B in the Appendix. The model revealed an effect of Minimal Pair Type. Discrimination performance was highest in trials where the different pair contained a non-minimal pair (91% accuracy) compared to trials containing voweled minimal pairs and vowelless minimal pairs (CCC vs. CCC: 80%; CVC vs. CVC: 77%). There was no difference between non-minimal pair trials and trials where the contrast was a voweled vs. vowelless minimal pair (CVC vs. CCC: 82%).

Figure 3

Experiment 1. Performance on paired discrimination trials across Minimal Pair and Speech Style conditions for native and non-native (English) Tashlhiyt listeners.

There was also an interaction between Minimal Pair Type and Speech Style. Post hoc Tukey tests were performed on the model with emmeans to unpack this interaction. First, the difference between discrimination performance for Casual and Clear speech styles was largest for the CVC vs. CVC comparison, and this difference was significant [est. = –1.6, z = –7.31, p < 0.001]. As seen in Figure 3, the lowest discrimination performance was for CVC vs. CVC trials in Casual speech. Post hoc tests also revealed that discrimination was worse in Casual speech for CCC vs. CCC trials [est. = –1.0, z = –6.4, p < 0.001]. However, there was no difference in discrimination across speech styles for non-minimal pairs [z = –2.3, p = 0.32] and for CVC vs. CCC trials [z = –2.8, p < 0.08]. Finally, a comparison between performance for non-minimal pair trials and CVC-CVC trials in Casual speech was significant, indicating that voweled trial discrimination was the lowest in reduced speech [est. = 1.7, z = 3.1, p < 0.05].

Language Background of the listener also predicted overall discrimination performance: Tashlhiyt listeners (87%) are better at discrimination of Tashlhiyt words than English-speaking listeners (77%). There was also an interaction between Language Background and Style: the positive coefficient value indicates that the difference between language backgrounds is greater in Clear speech than in Casual speech (indeed, the estimate for the post hoc comparison of non-native vs. native listeners for Clear speech [est. = –1.4, z = –4.6, p < 0.001] has a larger absolute value than that for the non-native vs. native Casual speech comparison [est. = –0.8, z = –3.9, p < 0.001]).

As mentioned above, the inclusion of the three-way interactions with Language Background and the other predictors did not significantly improve model fit. This indicates that, while English listeners overall perform lower on discrimination of Tashlhiyt words than native listeners, they do not display even lower performance of a function of any of the word or speech style types in the experiment.

2.2.2. Discrimination performance across sonority profiles

We also explore the extent to which sonority patterns within vowelless words lead to differences in perceptual sensitivity. To that end, two logistic regression models were run separately on the CVC vs. CCC trials subset and the CCC vs. CCC trials subset. Both models contained three fixed effects. The first two effects were the same across the two post-hoc models and coded identically as in the main model: Language Background (Tashlhiyt, English) and Style (Clear, Casual). The third fixed effect, Minimal Pair Type, was coded differently across the two models. In the CVC vs. CCC model, Minimal Pair Type contained 3 levels: CVC vs. CCC-Falling, CVC vs. CCC-Plateau, and CVC vs. CCC-Rising. In the CCC vs. CCC model, the three levels of the predictor Minimal Pair Type were CCC-Falling vs. CCC-Falling, CCC-Plateau vs. CCC-Plateau, and CCC-Rising vs. CCC-Rising. The random effects structure of both models was identical to that for the main model. As with the main model, we performed model comparisons to evaluate the inclusion of each fixed effect and all two- and three-way interactions. For the CVC vs. CCC model, only the model with simple fixed effects, excluding all possible interactions, was computed as the best fit and thus retained (model ANOVA Chisq = 17.5, df = 1, p < 0.001). For the CCC vs. CCC model, inclusion of interactions between Minimal Pair Type and Style and Language Background and Style resulted in the best fit (model ANOVA Chisq = 4.8, df = 1, p = .03).

The summary statistics for both models are provided in Tables C and D the Appendix. The CVC vs. CCC model revealed effects of Speech Style (higher discrimination overall for Clear than Casual speech). There was also an effect of Minimal Pair Type. Performance was lowest for the CVC vs. CCC-Rising trials (75%), where sonority profile makes it harder to discriminate, compared to the CVC vs. CCC-Falling (88%) and CVC vs. CCC-Plateau trials (86%) which contain larger sonority differences. Speech style did not mediate these differences.

Within vowelless trials, performance was lowest when both pairs had Falling sonority (71%), compared to when both pairs had Plateauing sonority (78%); trials containing vowelless words with different sonority (Rising vs. Plateau) had the highest performance (87%). For the CCC vs. CCC model, there was a significant interaction between Minimal Pair Type and Speech Style, which can be seen in Figure 4. Post hoc Tukey tests revealed that while discrimination was lower in Casual than Clear speech for Plateau [est. = –1.5, z = –5.8, p < 0.001] and Falling [est. = –0.8, z = –3.1, p < 0.05] CCC minimal pairs, there was no casual speech reduction for Rising vs. Plateau vowelless words [p = 0.33]. Thus, vowelless pairs with matching sonority are harder to discriminate in reduced speech.

Figure 4

Experiment 1. Performance on paired discrimination trials with vowelless words only across Sonority Profile and Speech Style conditions for native Tashlhiyt and English listeners.

2.2.2.1. Post hoc investigating role of transitional vocoids

How might the realization of a transitional vocoid affect listeners’ ability to discriminate vowelless words? We conducted a post hoc analysis to investigate this question. We tested whether the duration of the transitional vocoid present on the vowelless words in a different-pair affected listeners’ discrimination performance for trials that contained the same sonority profile (plateau v. plateau and falling v. falling, since performance was lowest for them within CCC trials). To that end, we ran a mixed effects logistic regression model on the discrimination data for the plateauing and falling CCC vs. CCC trials. The model included fixed effects of longest Transitional Vocoid Duration (in seconds, continuous variable, centered and scaled; a value of 0 was coded for the few trials where there was no transitional vocoid on any of the words) and Minimal Pair Type (Falling [reference level], Plateau), as well as their interaction. The model also included random intercepts for each participant, word pair type, and word pair ordering.

The model revealed a significant effect of Transitional Vocoid Duration [est. = 0.72, z = 2.5, p < 0.05] revealing that listeners perform better when the transitional element in a vowelless word is longer. There was also a significant interaction between Transitional Vocoid Duration and Minimal Pair Type indicating that longer vocoid duration provides an even larger boost for discrimination of plateauing sonority than for falling sonority vowelless words [est. = 1.04, z = 2.7, p < 0.01].

2.3 Interim discussion

In Experiment 1, native Tashlhiyt and non-native Tashlhiyt (English-speaking) listeners show variation in word discrimination performance for different types of minimal pair contrasts. As expected, minimal pairs are harder to discriminate between than non-minimal pairs. However, within minimal pair trials, perceptual sensitivity to vowelless minimal pairs (contrasting in the center segment) is not different overall than for trials with words containing vowels. Thus, our results indicate that vowelless words are not perceptually harder than words with vowels.

We also found that vowelless word pairs containing falling sonority profiles (e.g., nsˤħ vs. nʒħ) are harder to discriminate than pairs containing plateaued sonority (e.g., ʁbr vs. ʁdr). This result is surprising considering proposals that consonant sequences containing larger sonority differences will be easier to discriminate than those with smaller sonority differences (Ohala & Kawasaki-Fukumori 1997). However, the lowest performance for falling sonority contrasts is consistent with the observation that falling sonority clusters are the least common cross-linguistically and prior findings that they are most perceptually dispreferred by listeners from a range of language backgrounds (Berent et al. 2009).

Furthermore, a post hoc analysis revealed that listeners were better able to discriminate vowelless words when they contained longer transitional vocoids, and this was even stronger for plateauing sonority words. Thus, the phonetic realization of vowelless words in Tashlhiyt affects listeners’ ability to perceive their internal structure.

Notably, while non-native listeners performed overall lower than native Tashlhiyt listeners, the perceptual discrimination patterns within each contrast were parallel across listener groups. Thus, the lack of native language experience with Tashlhiyt does not make vowelless words less discriminable relative to other Tashlhiyt words. Relative perceptual sensitivity to vowelless words varying in sonority patterns was also similar across native and non-native listeners, further supporting that cross-language principles influence discriminability of vowelless words.

Additionally, vowelless words produced in Clear speech are better discriminated than when those words are produced in a more reduced speaking style. The biggest clear speech perceptual boost is for minimal pairs contrasting in vowels. This is not surprising since vowels are more informative than consonants. However, it is unexpected that trials containing CVC pairs in reduced speech had the lowest discrimination performance (hence, clear speech is actually enhancing the discriminability of voweled words to be as equally discriminable as other minimal pair trial types; in reduced speech, CVC minimal pairs are the hardest to discriminate). In the general discussion, we explore possible explanations for this finding.

We also observe that, within vowelless words, clear speech boosts discriminability of the harder sonority contrasts to a greater extent. In particular, falling and plateau sonority vowelless words are perceptually enhanced. Moreover, enhancement is observed for both listener groups. This adds to growing empirical work about how listener-oriented speaking styles can impact the perceptibility of a range of cross-linguistic phonological patterns (Smiljanić & Bradlow 2005; Kang & Guion 2008).

3.1. Methods

3.2. Results

We modeled nonword acceptability ratings of the 14 target nonword items as a continuous dependent variable (0–100; centered and scaled prior to model fitting) with a linear mixed effects model using the lme4 R package. The model included three fixed effects: Language Background (Tashlhiyt, English [reference level]; treatment coded), Speech Style (Clear, Casual [reference level]; treatment coded), and Sonority value of the middle segment. We adopt the sonority scale of Parker (2002) where integers are assigned to sounds based on their sonority ranking (vowel = 8, glide = 7, liquid = 6, nasal = 5, voiced fricative = 4, voiceless fricative = 3, voiced stop = 2, voiceless stop = 1). Sonority value of the middle segment was a continuous predictor, centered and scaled prior to model fitting. The random effects included by-participant and by-item random intercepts, as well as by-participant random slopes for Sonority Value and Speech Style (including by-participant random slopes for the interaction between Sonority Value and Speech Style resulted in a singularity error indicating over fitting).

Just as we did for Experiment 1, the addition of 2- and 3-way interactions between predictors was evaluated via model comparisons. The model with all two- and three-way interactions had the lowest AIC (model ANOVA Chisq = 20.4, df = 1, p < 0.01) and was thus retained.

The model output is provided in Table E in the Appendix. The model revealed an effect of Language Background wherein Tashlhiyt listeners provide higher acceptability ratings of the nonwords than English listeners (Tashlhiyt mean = 54.1, English mean = 36.1). There was also an effect of Speech Style: nonwords produced in Clear speech have higher ratings (Clear = 48.8, Casual = 41.4). There was also an interaction between Sonority value of the middle segment and Speech Style. In general, nonwords with a higher sonority middle segment are rated as more word-like if they are produced in Clear speech.

However, there was also a three-way interaction between listener Language Background, Sonority Value, and Speech Style. This interaction is depicted in Figure 5, which displays the mean nonword acceptability ratings by middle segment and Speaking Style, across Tashlhiyt and English-speaking listeners. As seen, English listeners are more likely to provide higher nonword acceptability ratings as the middle segment increases in Sonority – but the increase is only for items produced in Clear speech; acceptability ratings are low for phonetically reduced items across the board for the non-native Tashlhiyt listeners. Meanwhile, Tashlhiyt listeners display the reverse sonority-based acceptability pattern: nonwords containing middle segments that decrease in sonority are more likely to receive higher word-like judgements, especially when they are produced in Clear speech. Figure 6 provides the mean nonword acceptability ratings across speech styles and listener groups, but classifying the middle segment as either a vowel, sonorant, or obstruent. The interaction is more visually apparent in Figure 6: while English listeners give higher ratings to vowel and sonorant-nuclei items produced in Clear speech, Tashlhiyt listeners show preference for obstruent-centered words when they are produced in Clear speech.

Figure 5

Nonword acceptability ratings by Middle Segment Type, by Speech Style for Tashlhiyt and English listeners.

Figure 6

Nonword acceptability ratings where the middle segment is classified as a vowel, sonorant, or obstruent, by Speech Style for Tashlhiyt and English listeners.

4.1. Perception of vowelless words in Tashlhiyt

Vowelless words are rare in languages of the world. However, we find that they were not more difficult to discriminate in Tashlhiyt than words with vowels. This is surprising given that vowels provide robust transitional coarticulatory cues about the identity of adjacent sounds (Liberman et al. 1954) and that consonant clusters are harder to perceive than single consonant onsets (Foss & Gernsbacher 1983; Cutler et al. 1987). On the other hand, since Tashlhiyt has a large consonant inventory, there is pressure for speakers to enhance the acoustic cues to segment identity in vowelless words, even in casual speech where consonant cues are more likely to be obscured. Indeed, prior articulatory work has shown that consonant sequences are non-overlapping in Tashlhiyt and often produced with transitional vocoid elements that could provide acoustic cues to the segmental composition of vowelless words (e.g., Ridouane & Fougeron 2011). Indeed, the acoustic analysis of our stimuli revealed that vocoid-like elements, and even full schwas, are frequent in the vowelless items. Moreover, we found that they occur at similar rates across casual and clear speech, though they are longer in clear speech when they occur. And, we found that vowelless words containing longer transitional vocoid elements were easier to perceive. Thus, it appears that speakers do provide cues about the internal structure of vowelless words in Tashlhiyt. This parallels findings that word-initial stop+stop sequences in Tsou are produced with audible release bursts frequently and at similar rates across slow, normal, and fast speaking rates (Wright 1996). In our study, these cues are present even when producing reduced speech, though they are enhanced in speech produced with the purpose of making words more intelligible. Our present findings are consistent with a cue enhancement view that vowelless words in Tashlhiyt are produced in such a way so as to provide salient acoustic cues that help listeners recover segmental differences within them.

Furthermore, within vowelless words in Tashlhiyt, we found variation in perceptual patterns based on the sonority profile of consonant sequences. We observed that discrimination was harder when vowelless pairs contained matching word-onset sonority (here, falling vs. falling or plateau vs. plateau), than when sonority varied (rising vs. plateau). We interpret this as demonstrating that greater acoustic similarity across vowelless words made discrimination more difficult (Ohala & Kawasaki-Fukumori 1997). Moreover, discrimination performance for sonority-matching pairs was even lower in casual speech. Thus, reduced speech made it even harder to perceive acoustically similar vowelless words. However, in clear speech, matching plateauing sonority items were not harder to discriminate than when vowelless pairs contained different sonority profiles. There was also a boost for the matching falling sonority items. Thus, clear speech did provide a boost for hard-to-discriminate vowelless contrasts.

Our acoustic analysis suggests that the length of the transitional elements is enhanced in clear speech. A post hoc analysis also established a relationship between the perceptual patterns and the realization of schwa across these word types: vowelless words are easier to discriminate when they contain transitional vocoids that are longer in duration. This provides a direct link between the phonetic properties of vowelless words to their perceptual stability. This also adds to cross-linguistic evidence that clear speech provides more auditorily robust phonetic variants of words (e.g., Smiljanić & Bradow 2005), extending this to vowelless words in Tashlhiyt.

There are additional factors though, not considered in the present study, that might explain discrimination, or general perceptual, ability for vowelless words in Tashlhiyt. For instance, here, we focused only on the initial two segments of vowelless words. Yet, the phonological and phonetic properties of the word-final segment could impact phonetic and perceptual properties of vowelless words. Future work exploring the effect of the final consonant can be a fruitful direction. Moreover, the segmental composition of words has also been shown to predict vocoid realization. For instance, prior work has shown that schwa-like elements are common before final /r/ (Ridouane & Cooper-Leavitt 2019); many of our lexical items had a final /r/, therefore this could explain why some of our items with falling sonority were difficult to discriminate when they both contained a final /r/ (e.g., nkr vs. ngr). It could also explain why the rising vs. plateau comparisons were easier to discriminate when one pair contained an /r/ (e.g., kʃf vs. krf). Another consideration that we did not explore in the current study is that the laryngeal properties of adjacent consonants is another systematic factor conditioning the realization of vocoids within vowelless words. For instance, prior work has found that the occurrence of schwa-like elements relates to the voicing and place of articulation of adjacent consonants (e.g., Ridouane & Fougeron 2011). Also, pharyngealization of a segment within a vowelless word might influence its perceptibility (e.g., ftħ vs. frˁħ). Thus, there are indeed many additional factors that systematically predict the presence and duration of vocoid-like elements in vowelless words. These present ripe avenues for future work to explore their influence on the perception of these words in Tashlhiyt. Additionally, our classification of words in different sonority categories was based on a binary ranking of consonants (O vs. S). Work employing more fine-grained classifications of sonority could provide even further insights into how acoustic patterns of different segment types influence perceptual ability of Tashlhiyt vowelless words.

4.2. Perception of words with vowels in Tashlhiyt

We also observed that words containing vowels received the lowest discrimination performance in reduced speech. Why are voweled words in Tashlhiyt more difficult to perceive than vowelless words in this condition? One possibility is that strong consonant-to-vowel coarticulatory influences in Tashlhiyt leads to substantial variation across vowels in CVC words in reduced speech such that discriminating between different vowels is difficult. There is some work suggesting that consonant-to-vowel coarticulation leads to a large amount of phonetic variation in vowels for languages with a high consonant-to-vowel ratio, like Tashlhiyt. For example, it has been shown that there is substantial acoustic variation in vowels based on consonantal context in Arabic (also a language with a 3-vowel vs. many consonant inventory) (e.g., Embarki et al. 2007; Bouferroum & Boudraa 2015). High vowel variability due to consonant-to-vowel coarticulation has been reported for other languages with larger consonant-to-vowel ratios (e.g., Salish in Bessell (1998), Yanyuwa and Yindjibarndi in Tabain & Butcher (1999)).

Notably, the difference in perceptual performance across voweled and vowelless words that we observed in Tashlhiyt indicates greater consonant contrast preservation than vowel contrast preservation. This can be explained by the gestural coordination patterns in the language: consonant-to-consonant coarticulation is minimal, but there is extensive consonant-to-vowel coarticulation. In effect, then, vowels in Tashlhiyt are not as acoustically informative as consonants for lexical discrimination. Note, however, that the discrimination conditions containing voweled-vowelless minimal pairs (CVC vs. CCC) were equivalent in performance with the non-minimal pair trials, indicating the voweled and vowelless words were distinct enough from each other to be easy to distinguish perceptually. Within those trial types, words with similar acoustic profiles (CVC vs. CCC-Rising) were more difficult to discriminate in both reduced and clear speech, and those with the greatest acoustic difference in their sonority profile (e.g., CVC vs. CCC-Plateau) had higher performance across speech styles. Accordingly, Tashlhiyt vowels contrasting with a (obstruent) consonant were easier to discriminate, as opposed to vowels contrasting with sonorants and other vowels, which were hard to discriminate. Work exploring the production and perception of vowel variation in languages with high consonant-to-vowel ratios is a ripe avenue for future research.

4.3. Comparison of native and non-native listeners

We also compared native and non-native groups of listeners. Tashlhiyt listeners displayed overall higher discrimination performance than English listeners; however, the relative patterns across word types were basically the same for the two listeners groups. Thus, discrimination of vowelless words in Tashlhiyt, in particular, does not appear to be modulated by experience with the language. In contrast, there were robust differences in Tashlhiyt nonword acceptability ratings across listener groups. Tashlhiyt listeners were more accepting of vowelless nonwords with an obstruent center, while English listeners preferred nonwords with vowel nuclei. These cross-language differences can be related to listener native language-experience: vowelless words with obstruent centers are the most common structure in Tashlhiyt CCC words (around 32% of tri-segmental verbs in Lahrouchi 2010), so the obstruent-centered nonwords in Experiment 2 resemble many real words in Tashlhiyt. We also acknowledge that our liquid-centered nonwords actually violate Tashlhiyt phonotactics of having two adjacent liquids. Indeed, those forms received the lowest word-like ratings which is further support that listeners were recruiting their lexical knowledge (i.e., vowelless words with two adjacent liquids are not in their lexicon) when completing this task. Meanwhile, English does not allow vowelless words (a preference for sonorant-centered over obstruent-centered nonwords by English listeners perhaps also reflects the allowance of syllabic sonorants in English, even though this is permitted only in unstressed syllables).

We also found that the clear speech benefit in the discrimination task was larger for Tashlhiyt listeners than for English-speaking participants. This aligns with prior findings that clear speech enhances intelligibility for native speakers to a greater extent than non-native listeners (Bradlow & Bent 2002). Also notable is that the largest cross-group differences in the ratings task were for the clear speech items. This is the first study, to our knowledge, comparing word-likeness ratings across items produced in different speaking styles. Presumably, the clear speech forms were easier to parse, phonologically, and thus provided more intelligible words that allowed for phonotactic preferences to emerge. In this way, while some paradigms might make clear speech appear to be “native-listener oriented” (cf. Bradlow & Bent 2002), we find there are some tasks where the perceptual benefit of clear speech is greater for non-native listeners; here, for the rating of voweled nonwords as being more like a word of English.

More broadly, that we observed overall qualitative differences in the native and non-native listeners’ performance across perception tasks can also speak to the phonological processing of vowelless words. Discrimination relies more on low-level, auditory mechanisms and differences within vowelless words are paralleled across the listener groups. Thus, besides the overall greater perceptual sensitivity by native listeners, experience with Tashlhiyt did not qualitatively change listeners’ ability to discriminate when the word pairs were acoustically distinct enough. However, for the word-likeness task, performance recruits higher-level phonological processing mechanisms, and native-language-specific phonotactic patterns influenced the patterns across listener groups. Thus, the perception of vowelless words has both a cross-linguistic, auditorily-based processing mechanism and a phonological component that recruits native-language-specific phonotactic patterns.

4.4. Implications for phonological typology

If vowelless words are not categorically difficult to discriminate, why are they so cross-linguistically rare? Articulatory and acquisitional influences have also argued to shape the evolution of consonant sequencing preferences (Ohala & Kawasaki-Fukumori 1997; Blevins 2004). For example, children’s first words are typically CV(C), despite hearing words that contain more complex structures (de Boysson-Bardies 1999; McLeod, Doorn, & Reed 2001). For articulation, speech production models propose that CV sequences have a unique coordination relationship with one another, which makes that structure particularly stable (Saltzman & Byrd 2000; Tilsen 2016). Also, adjacent consonants with very different places of articulation, for instance, might be articulatorily difficult if the tongue configurations are incompatible (Ohala & Kawasaki-Fukumori 1997). Therefore, articulatory biases could explain why vowelless words are not more common cross-linguistically.

Moreover, as mentioned above, Tashlhiyt has a unique phonetic profile - wide, non-overlapping consonant sequences with frequent audible, vocoid-like elements - which make consonant sequences recoverable to listeners. Noncoarticulating consonant clusters don’t appear to be cross-linguistically uncommon, but complex clusters with heavy overlap appear to be more frequent across languages where the articulatory dynamics of consonant sequences has been studied (Hermes et al. 2017). It is possible that the phonetic pre-conditions of non-overlapping consonants with audible transitions permit the maintenance of vowelless words in languages where they are allowed. Language-specific phonetic implementation of sound sequences could explain the synchronic stability of complex phonotactic structures in Tashlhiyt. These speculations open lines of inquiry that can be explored in future work. For instance, future work directly comparing the perception of consonant clusters across languages that are more coarticulating (e.g., English) vs. less coarticulating (like Tashlhiyt) could provide further insight into how cross-linguistic differences in articulatory dynamics might make some sound contrasts more or less stable. Moreover, while we used only naturally-produced tokens in the present study, manipulation of the phonetic properties of the vocoid elements within Tashlhiyt can be a further testing ground for understanding the relationship between phonetic variation and perceptual sensitivity of vowelless words.