1.1 Coarticulation and perceptually-driven sound change

In addition to its role in spoken word comprehension, coarticulation has been at the forefront of theoretical and empirical understandings about the relationship between synchronic phonetic variation and the evolution of phonological patterns in a language over time. For instance, the reanalysis of coarticulation as contrastive nasalization (e.g., VN > ṼN > Ṽ) is recognized as a common historical pathway of sound change (Chen & Wang 1975; Hajek 1997). How do coarticulatory-based sound changes actuate? Understanding the conditions and catalysts for coarticulatory-based sound changes, as well as the pathway and time course by which gradual phonetic shifts occur over time within a speech community, has been the topic of much theoretical work (Weinrich et al. 1968; Beddor 2009; Yu 2023). This is because coarticulation results in an acoustic signal that contains temporally-distributed phonetic cues. For example, a speech signal containing an utterance [ṼN] could be interpreted by a listener as underlyingly /VN/, /ṼN/, or just /Ṽ/. A listener’s processing of the coarticulatory features, parsing the speech signal into discrete gestures, is hypothesized to be a source of innovative phonetic variants and a starting point for historical change in a language (Ohala 1993; Beddor 2009; Harrington 2012).

One view is that listeners perceive the speech signal as a consequence of co-produced articulatory gestures (Fowler 2005). Thus, listeners perceive coarticulation by attributing the acoustic effects of gestural overlap in the speech signal to their articulatory source(s) (Fowler 1984). Listeners have been shown to display behavior indicating they factor out phonetic variation when they can ascribe that information to another source in the signal. For example, listeners’ category boundary of a /da-ga/ continuum shifts toward more alveolar responses when the stimuli are presented in the context of a preceding /ɹ/, suggesting that the perceptual system attributes a lower F3 to the preceding rhotic (Mann 1980; also Mann & Repp 1980 for the influence of vowel coarticulation on sibilant perception). For nasal coarticulation, studies have demonstrated that listeners hear a nasalized vowel in the context of a nasal consonant as phonologically equivalent to oral vowels (e.g., [ṼN] is interpreted by the listener as /VN/, Kawasaki 1986). In this scenario, there is “parity” between the production and perception of coarticulation. This is a phonetically stable condition in that phonological parity has occurred between speaker and listener, a situation that would not lead to phonological change.

However, theoretical models of listener-based diachronic instability propose that since coarticulation results in an ambiguous acoustic signal, sound change can be actuated when listeners and speakers come to different interpretations of an utterance. Ohala’s (1993, inter alia) “innocent misperception” model is one hypothesis for how listener-based coarticulatory processing mechanisms lead to novel percepts due to coarticulation. For instance, if the listener does not fully attribute the coarticulatory variation to its source (e.g., a coarticulatorily nasalized vowel could be analyzed by the listener as phonologically nasal), this can lead to the decontextualization of coarticulation as intrinsic to the vowel. Such a scenario has been argued to be particularly likely if a listener fails to detect the source of coarticulation, such as if the nasal consonant was deleted or obscured in the speech signal for some reason (such as by noise) or if the listener makes a decision about the phonological structure of the vowel before even hearing the final consonant (Beddor 2009; Beddor et al. 2013).

In Ohala’s model, these listener-driven sound change mechanisms are framed as “errorful” (i.e., a listener’s failure to fully correct). In contrast, another perspective is that perceptually-driven sound changes reflect differences in the relative weighting of coarticulatory cues in the speech signal across listeners (e.g., Beddor 2009). Under such a view, coarticulatory-based sound changes are gradient and could reflect either differences in strategies or distinct coarticulatory grammars across individuals. For instance, Beddor (2009) has found that listeners vary in their use of nasal coarticulation in perception; for some US English listeners, vowel nasalization alone has been found to be sufficient to determine that a speaker said “Ben” (rather than “bed”) whereas other listeners require a nasal consonant. She proposes that listeners who are innovative with respect to perceived coarticulation are those who attend more closely to the coarticulatory information and assign less weight to—and perhaps do not require—the coarticulatory source. In this approach, the impetus of sound change is not misperception but rather incremental differences in the relative perceptual importance or weight of the multiple phonetic properties associated with a phonological contrast that occur across individuals.

Such cue-based models of perceptually-driven sound change argue that phonologization of coarticulation occurs along a pathway where, for instance, there is first a contextual relationship between coarticulation and source, but then listeners’ perceptual attention shifts gradually to the presence of the coarticulatory feature alone – whether or not the source is present – as the dominant cue to a contrast (Beddor 2009). This can occur over time as a gradual trade-off between the temporal extent of coarticulation and the strength of the source (Beddor et al. 2007). A final stage occurs when the cue is realized in the complete absence of the source. Beddor et al. (2013) demonstrate this with eye tracking: during the time course of the vowel alone, US English listeners are more likely to look to images depicting the CVN(C) word when presented with enhanced anticipatory nasal coarticulation, than when there was a vowel containing less coarticulation.

Many empirical studies of speech production have found gradient differences in overall degree and extent of coarticulatory nasality across languages; this is consistent with the view that coarticulatory-based sound change involves cue-reweighting mechanisms rather than categorical changes in the status of vowel nasality. For instance, there are weaker nasal coarticulatory patterns in CVN words in Spanish, French, Swedish, Italian, Japanese and Bininj Kunwok relative to that observed in other languages (Ushijima & Hirose 1974; Clumeck 1976; Cohn 1990; Farnetani 1990; Solé 1992; Delvaux et al. 2008; Stoakes et al. 2020). Meanwhile, US English, Sundanese, and Tereno have extensive coarticulatory vowel nasalization, with the velum lowering gesture starting very early in vowels before the nasal coda (Cohn 1990; Ohala 1993). Evidence of cross-dialectal differences in nasal coarticulation within a language indicates that coarticulatory patterns can vary in gradient ways at the speech community level, further supporting the cue-reweighting framework for coarticulatorily-motivated sound change. For instance, Bongiovanni (2021) found that extent of anticipatory vowel nasalization differs across dialects of Spanish: in Dominican Spanish, coarticulation is extensive and “phonologized”, while Argentinian Spanish has less extensive nasalization. These findings are consistent with related work by Solé (2007) suggesting gestural overlap in continental Spanish was a mechanical product of articulation and not a linguistically-specified feature (i.e., /VN/). In a comparison across varieties of Afrikaans, Coetzee et al. (2022) found that Kleurling Afrikaans speakers produce less extensive anticipatory nasal coarticulation than white Afrikaans speakers. They also found that individuals — in both communities — who produced more extensive coarticulatory nasality relied more on nasalization during lexical perception, suggesting consistency of cue weighting across a speaker’s production and perception. Patterns of coarticulation have also been shown to vary across dialects of US English: speakers from Philadelphia, Pennsylvania produce more extensive nasal coarticulation than speakers from Columbus, Ohio (Tamminga & Zellou 2015). Even within a speech community, variation in nasal coarticulation across generations has been observed. For instance, Zellou and Tamminga (2014) observed that Philadelphia English speakers born between 1950 and 1965 display a trend of increasing coarticulatory vowel nasalization, yet speakers born after 1965 move towards decreasing coarticulatory nasality; finally, speakers born after 1980 reverse this change, moving again toward producing greater nasal coarticulation. Taken together, these studies show that coarticulatory patterns change in gradient ways within a dialect (see also Carignan et al. 2021), and can even reverse direction.

The current study examines changes within California English, a US English dialect. The most well-documented sound change within California is the “California Vowel Shift”, consisting of the fronting of back vowels and a counterclockwise shifting of the front and low vowels over apparent-time (Eckert 2008). Less well known, but becoming increasingly investigated, is the change in the realization of vowels before nasal coda in California English, involving changes not just in degree of anticipatory nasalization, but also changes in diphthongization, vowel positioning, and other features across speech communities (Brotherton et al. 2019; Carignan & Zellou 2023), as well as over time (Hall-Lew et al. 2015; Podesva et al. 2015), and even across individuals within one social cohort (Zellou & Brotherton 2021).

Under the cue-reweighting framework of listener-based phonetic innovation (i.e., gradual change in the use of an acoustic feature as signaling a lexical contrast), several studies have investigated how the perceptual weighting of coarticulatory features changes across speech communities within a language as the catalyst for phonological change. Kleber et al. (2012) examined /ʊ/-fronting in Standard Southern British English and found younger (ages 18–20 years) and older (ages 50+ years) speakers had distinct perception grammars: younger listeners compensated less for alveolar consonant coarticulation on fronted /ʊ/ than older speakers. These group-level differences were argued to reflect different stages of a sound change: the older listeners reflect a pre-sound change stage where coarticulation is phonetic, while the younger listeners are in a stage where coarticulation has become phonologized. These perceptual differences were argued to be the catalyst for sound changes observed in production whereby older speakers displayed more coarticulatory-induced fronting of /ʊ/ while younger speakers show more categorical /ʊ/-fronting effects (i.e., fronting in both coarticulatory and non-coarticulatory contexts). Thus, the change in perceptual use of coarticulation over time was argued to be a trigger for phonological change in Standard Southern British English.

1.2 Multidimensionality of nasal coarticulation

Nasal coarticulation is a particularly apt phenomenon to examine in exploring the perceptual mechanisms for how sound change actuates. As mentioned, vowel nasalization results in variation across multiple temporal and spectral dimensions. For one, gestural overlap in the velum lowering gesture from a nasal coda onto a preceding vowel could be reanalyzed by a listener as an intrinsically nasalized vowel. Spectrally, this has been shown to be realized as an increase in the amplitude of the “nasal formants” and a dampening of the amplitudes of the “oral formants”, as well as an increase in the width of the formant band, especially for F1 (Chen 1997; Styler 2017). Since the acoustic characteristics of an intrinsically nasalized vowel (e.g., in the case of phonological nasalization, such as French) and a contextually nasalized vowel (i.e., coarticulated) are similar, a listener hearing a nasalized vowel does not know which one it is (unless a source consonant is present in the speech signal and detected).

Additionally, the spectral effects of vowel nasalization are similar to the spectral effects of breathy phonation (Ohala 1974; Matisoff 1975). Specifically, breathiness results in a weakening of the amplitude of the first formant along with an increase in the amplitude of the first harmonic (Garellek et al. 2016). Even though the increased spectral tilt in breathy vowels and nasal vowels stem from different articulatory sources (subglottal vs. velopharyngeal; Ohala 1974; Garellek 2014; Carignan 2017), the acoustic similarity between these features can lead to ambiguity for the listener. Indeed, breathiness on vowels has been reanalyzed by listeners as nasalization (Ohala 1975; Imatomi 2005) and, likewise, nasalization on vowels as breathiness (Blevins & Garrett 1993; Ohala & Busà 1995; Blevins 2004; Garellek et al. 2016).

Moreover, formant frequency patterns can also be perceptually perturbed by the presence of acoustic nasalization on vowels that can lead to sound change (Carignan 2014). Nasalization shifts the perceptual center of gravity of the spectrum lower for non-high vowels. A perceived lower F1 can be reanalyzed as due to changes in tongue position for these vowels (Beddor, Krakow, & Goldstein 1986; Krakow, Beddor, & Goldstein 1988; Carignan 2017). Most often, since nasal coarticulation is temporally dynamic, affecting portions of the vowel closer to a nasal consonant, the perceptual influence of nasalization on vowel quality changes result in the reanalysis of a monophthongal vowel as diphthongal. This can explain many nasality-induced diphthongal effects in vowel systems, cross-linguistically (Parkinson 1983; Walker 1984; Baker et al. 2008; Mielke et al. 2017; Zellou & Brotherton 2021).

1.3 Current Study

The aim of the current study is to investigate whether the perceptual use of nasal coarticulation changes over apparent-time in California English. We aim to explore whether there are changes in listeners’ perceptual weighting of cues for nasal coarticulation across apparent-time with California English listeners ranging in age from 18 to 58.

We compared perception of vowels in CVN and CVC words produced by speaker groups whose vowels are either non-ambiguous to listeners (‘Extensive Coarticulators’) or generate ambiguity (‘Minimal articulators’) in a prior study (Zellou 2022). This is analogous to the manipulated extensive and minimal coarticulation conditions in prior work (Beddor et al. 2013). Here, our stimuli are non-manipulated speech to reflect naturally occurring variations across several individuals. Moreover, our task is a lexical completion task, similar to that used in prior work (Ali et al. 1971; Ohala & Ohala 1995), which excises the final coda. This is analogous to situations in natural interactions where the coda is obscured (i.e., by noise) or deleted (via hypoarticulation). Excising the coda also allows us to test how our listeners respond in the stage of phonologization where perceptual attention is shifting more toward coarticulation, and away from the source. This approach probes listeners’ use of acoustic cues present on vowels that signal the phonological status of the final consonant as being nasal or oral.

3.1 Perceptual responses by listener age across vowel and speaker types

The model output is provided in Table 3 and the summarized raw data is plotted in Figure 2. First, the model showed an effect of Structure: overall, listeners had lower accuracy on average in identifying stimuli originally from CVN words. Additionally, there was an effect of Coarticulation Category: overall, listeners had lower accuracy in identifying the original lexical item from vowels produced by the Minimal Coarticulator speakers. However, this was mediated by a two-way interaction between Structure and Coarticulation Category: listeners were worse at lexical identification for the coarticulated vowels from CVNs produced by the Minimal Coarticulators.

Figure 2

Mean proportion of trials where participants identified the originally intended CVC or CVN word based on hearing the excised CV syllable, by listener age group (younger = 18–25 years old “Generation Z”, middle = 26–41 years old “Millennial”, older = 42–58 years old “Generation X”). The colors show differences for the speaker types: those who produced a higher degree of nasal coarticulation (“Extensive Coarticulators”; in light grey) and those who produced a lesser degree of coarticulation (“Minimal Coarticulators”; in dark green). Error bars indicate standard error of the mean.

There were also three-way interactions between Structure, Coarticulation Category, and Listener Age Group⁵, which can be seen in Figure 2. As seen, performance for the Minimal Coarticulators’ nasalized vowels is higher for younger listeners but lower for older listeners. The releveled model (provided in Supplementary Material, Table S1) showed no effects for the middle group, nor any interactions for the middle group with Structure or Coarticulation Category.

No other effects or interactions were observed. In particular, we do not see evidence of an overall Listener Age Group effect, such that participants’ performance on the task overall does not appear to worsen with age; rather, older listeners display reduced ability to identify the originally intended word from the coarticulated vowels produced by the Minimal Coarticulators specifically, relative to higher performance of these tokens by the younger listener group.

3.2 Acoustic cue weighting by listener age for coarticulated vowels

Next, we tested whether there is variation in the use of the multiple acoustic cues to vowel nasalization during perception across the listener groups. We fit a mixed effects logistic regression model to the responses for only the nasalized vowels (i.e., vowels from CVNs) with the lme4 R package (Bates et al. 2014). The model included fixed effects of Listener Age (3 levels: older, middle, younger listener age groups), Acoustic Nasality (A1-P0), F1 Bandwidth, and Diphthongization (ROC)⁶. In addition, interactions between Listener Age (younger, middle, older listener age groups) and each acoustic variable were included in the model⁷. The model also included by-Listener, by-Item, and by-Speaker random intercepts. We additionally included by-Listener random slopes for each of the acoustic features (Acoustic Nasality, F1 Bandwidth, Diphthongization)⁸. Acoustic variables were standardized around the mean using the scale() function in R prior to model fitting, while factors were sum coded.

While collinearity could be a problem when running a model with multiple acoustic features correlated with a phonetic cue, a check for multicollinearity using the check_collinearity() function in the performance R package (Lüdecke et al. 2021) computed a ‘low correlation’ assessment. The variance inflation factor (VIF) values for all the predictors were below 1.5, indicating low collinearity among variables and thus all were retained in the model.

The summary statistics for the model are provided in Table 4 (the releveled model output showing the middle listener age group is provided in Supplementary Material, Table S2). In terms of overall acoustic cue use, all listeners used larger F1 bandwidth and higher spectral rate of change (ROC) in formants to correctly identify coarticulated vowels as originating from CVN words, but there was no overall effect of A1-P0 (acoustic nasality).

Additionally, Listener Age Group interacted with two of the acoustic variables. First, there was an interaction between Listener Age Group and A1-P0 (acoustic nasality). Since acoustic nasality increases with decreasing A1-P0 values, we see that younger adult listeners show a stronger reliance on acoustic nasality in identifying CVN items from coarticulated vowels. The middle listeners also showed an interaction with A1-P0, with weaker reliance on the cue (recall that nasality is indicated with a negative sign) [Coef = 0.24, z = 2.75, p <0.01]. Yet, there was no difference in reliance on acoustic nasality for older adults. There were also interactions between Listener Age and F1 bandwidth, wherein older listeners use F1 bandwidth as a stronger cue to nasalization, while younger listeners use F1 bandwidth less. The releveled model showed that there was no difference for the middle listener age group in use of F1 bandwidth. Finally, there were no interactions observed between Listener Age and ROC suggesting that use of diphthongization as a cue to nasality remains stable across apparent-time.

4.1 Evidence for sound change being perception- or production-led?

Prominent models of sound change argue that innovation in either the production or perception of coarticulation triggers phonological change at the community level (Harrington 2012). Thus, a question that has garnered much attention in the literature is whether coarticulatorily-motivated sound change propagation is perception- or production-led (Kleber 2018; Kuang & Cui 2018; Coetzee et al. 2018; Pinget et al. 2020). One possibility is that sound change first propagates through shifts in perceptual attention, then is later realized in speakers’ productions; that is, sound change propagation is perception-led (Harrington 2012). Evidence of this comes from apparent-time studies showing, for instance, that younger speakers perceptually attend to coarticulatory patterns that they do not yet produce (e.g., Kuang & Cui 2018 find that some younger Southern Yi speakers who do not produce the more innovative phonetic shift still attend to it in perception). Another possible scenario is that sound change is propagated when coarticulatory effects are exaggerated by listener-turned-speakers and then get reanalyzed as the primary feature; that is, sound change propagation is production-led (Hyman 1976; Baker et al. 2011). This was observed by Pinget et al. (2020) who found that individuals who were most advanced in obstruent devoicing in Dutch were still perceptually sensitive to the conservative pattern.

Is there a concomitant, parallel change in production in this speech community? If yes, is this a production- or perception-led sound change? Examining any changes in the production of the multiple features associated with nasal coarticulation across time in California English can be informative as to whether California speakers are indeed moving more toward greater use of acoustic nasality as a primary phonetic cue to this contrast, and using F1 bandwidth less, as we would predict based on the perception findings in the current study; indeed, prior studies have shown that acoustic nasality and F1 bandwidth are features modified by speakers across contexts, and are therefore under active speaker control (e.g., Lu & Cooke 2008; Zellou & Scarborough 2015; Zellou & Scarborough 2019). Examination of how the timing of such a change in production is aligned with our perceptual shift can directly speak to how misalignment triggers phonological change over time in a speech community.

At this time, however, we can only speculate. Do we have any evidence for what is happening in production? In the present study, our stimuli came from multiple speakers in the same, younger age group (18–20 years old). So, if there is a change-in-progress in production, it is not strictly age-graded; younger speakers produce both “innovative” and “conservative” patterns in signaling the CVC-CVN contrast on vowels. This within-age category variation is consistent with cross-speaker differences reported by Zellou and Brotherton (2021) who found that some 19 year-old speakers were more advanced in the pre-nasal short-a split in California English, while others were more conservative. If we take the older listeners’ perceptual sensitivity in the present study as reflecting the conservative pattern, one possibility is that the Minimal Coarticulators represent a phonetic innovation in the reweighting of cues for nasalization cues away from F1 bandwidth. In this scenario, younger listeners show perceptual sensitivity to both the “conservative” pattern (Extensive coarticulation) and the “innovative” pattern (Minimal coarticulation), while older speakers are sensitive only to the “conservative” pattern (Extensive coarticulation). Thus, there are younger speakers who produce distinct coarticulatory features and younger listeners are also sensitive to both the innovative and the conservative patterns. The presence of exclusively innovative and conservative younger speakers — but perceptual flexibility for all younger listeners — could be evidence that this change is production-led. In Pinget et al. (2020), more innovative producers in Dutch maintain perceptual flexibility in order to understand more conservative speakers (similar patterns are also observed for Chru for some speakers in Brunelle et al. 2020). In California English, Zellou and Brotherton (2021) found that more advanced speakers in the ban~bad split still imitate (thus, are perceptually sensitive to) other speakers’ conservative nasality patterns. Therefore, the age-related changes in the current study could reflect a production-led shift, but nevertheless still one that is consistent with cue reweighting accounts of coarticulatory-based sound changes (Beddor 2009; Kuang & Cui 2018; Harrington et al. 2016). Future work exploring individual speakers’ production and perception patterns is needed to test whether these shifts might be perception- or production-led.

4.2 Future directions

In addition to exploring the production-perception link, our findings raise several other directions for future work. While other studies use methods that track listener behavior over time in clear and pristine listening conditions (e.g., eye-tracking in silence, Beddor et al. 2013), the current study more closely parallels natural language situations in the real world where part of the speech signal is obscured (by noise, or a cough) or deleted (as in hypoarticulation). Some researchers argue that sound change is more likely to be actuated in hypoarticulated speech contexts; when the source of coarticulation is reduced or obscured, listeners are less likely to attribute coarticulation to its source and instead can perceive it as an intrinsic property of the segment (Lindblom et al. 1995; Harrington et al. 2019). From the present study, we can deduce how listeners will react in such scenarios (e.g., background noise, hypoarticulated speech). A future direction is to compare how listeners attend to acoustic cues across different contexts such as with different types of noise and feedback from interlocutors, and also to coarticulatory patterns across systematic types of speech styles (e.g., nasal coarticulation can vary systematically across casual, slow-clear, and fast-clear speech in Cohn & Zellou 2023).

Future work can also compare the perceptual patterns in the present study with those by listeners from other dialect backgrounds or language experiences. For instance, is this a change in apparent-time for multiple dialects of US English, or just California English? Another avenue is to explore how listeners learn and adapt to novel cue weightings in production, given exposure and/or lexically-guided training with speakers such as our Minimal Coarticulators. Do listeners adapt their cue weighting of acoustic cues in using coarticulation for lexical perception? Furthermore, the type of speech produced — read speech — might have shaped the degree of style shifting. For example, related work has shown that speakers produce more variation in creakiness in some styles (e.g., conversational speech) than others (e.g., read speech) (Becker, Khan, & Zimman 2022), suggesting that we might see greater variation in more naturalistic contexts. We also acknowledge that our acoustic analysis involved taking acoustic measurements taken over several time points within a vowel and then transforming them into a single value for the statistical analyses. There is much work showing how coarticulation results in dynamic information (Carignan 2021), the time course of which listeners are exquisitely sensitive to (Beddor et al. 2013). Future work examining how cues unfold over time could reveal more insight into the dynamics of cue-use, both within a segment (e.g., in a vowel) but also across time (e.g., in a speech community). Finally, the speakers (Minimal and Extensive Coarticulators) were all California English speaking undergraduates (aged 18–20); future work examining how perception of coarticulation by age is needed, as well as the potential top-down effects of age-based expectations in shaping perception of coarticulation (e.g., /u/ fronting in Zellou, Cohn, & Block 2020).