Under listener-oriented models of sound change, perceptibility is argued to be one of the key drivers of phonetic change and thereby phonological typology (Blevins, 2004, 2006; Ohala, 1981, 1993). Acoustic ambiguity arises not only due to environmental noise that hinders perception, but also as a result of coarticulation, interspeaker variation (both physiological and sociolinguistic), and other factors. Ohala (1993) argues that listeners are typically able to correct for variability in the acoustic signal when its sources are predictable. A wealth of experimental evidence supports this conclusion and shows that listeners shift their perceptual boundaries according to phonological context (Mann & Repp, 1980, 1981) or their interlocutor (Hay, Nolan, & Drager, 2006; Hay, Warren, & Drager, 2006; Johnson et al., 1999). If the underlying source of variation becomes no longer predictable, such as if a coarticulation-triggering segment is itself misperceived or lenited, misperception-based models of change posit that listeners will re-interpret coarticulatory variability to be an inherent quality of the affected segment. Listeners who have re-interpreted a segment’s underlying form adopt production targets which differ from their predecessors, thereby initiating sound change.

This mechanism of change is clearly demonstrated by the development of nasal and nasalized vowels in languages like French and English, among many others (Beddor et al., 1986; Krakow et al., 1988). In English, coarticulatory vowel nasalization is pervasive and occurs due to anticipatory velar lowering before nasal codas. Nasalization of the preceding vowel alters its acoustic quality, particularly F1, which listeners may misattribute to the vowel’s oral articulation, namely a difference in tongue height. Krakow et al. show that American English listeners typically correct for this effect and accurately perceive the height of nasalized vowels, but only in contexts with nasal consonants (e.g., [bṼn] but not [bṼd]). In French, where vowel nasality is phonologically contrastive, the nasal vowel system exhibits a smaller number of height distinctions than the oral vowel system and the nasal vowels differ in height relative to the oral vowels from which they developed (Beddor et al., 1986).

In the case of nasal vowels, the articulators involved are presumably (mis)perceived auditorily, as configurations of the tongue, velum, pharynx, and larynx are mostly invisible. It is therefore unsurprising that nasal vowels exhibit a wide array of articulatory configurations, which emerge due to their acoustic complexity. When compared to their oral counterparts, nasal vowels may show differences not only in tongue height, but also in lip aperture, voice quality, and pharyngeal constriction (Carignan, 2014, 2017; Carignan et al., 2015; Chen, 2022; Chen & Havenhill, 2023; Garellek et al., 2016; Shosted et al., 2012). Some of these effects are argued to enhance the acoustic quality of nasalization (Garellek et al., 2016) or to preserve phonological contrasts (Carignan, 2018a; Chen, 2022), although interspeaker articulatory differences may also emerge due to misperception of the speaker’s articulatory configuration (Carignan, 2018b). Covert interspeaker articulatory variability is also notably observed for English /ɹ/ and /s/, which may be produced using a range of lingual gestures that yield acoustically similar output (e.g., Bladon & Nolan, 1977; Dart, 1998; Delattre & Freeman, 1968; Espy-Wilson, 2004; Mielke et al., 2016). This type of variation is well predicted by auditory models of speech perception and theories of acoustic enhancement. If speech sounds are optimized for their auditory-acoustic quality, then a speaker’s articulatory strategy is somewhat arbitrary as long as it yields the correct acoustic output (Johnson et al., 1993).

By contrast, configurations of the lips can be perceived not only auditorily, but also visually. Given this potential perceptual advantage, it is reasonable to hypothesize that visual perception of lip gestures may inhibit misperception-based change by disambiguating the speech signal. In turn, languages may be more likely to preserve contrasts that are auditorily and visually distinct, as opposed to those that can be distinguished through audition alone. For instance, in comparison with coronals and velars, labial sounds appear to be less likely to undergo changes to their place of articulation: debuccalization or full palatalization of labials are cross-linguistically rare changes (Bateman, 2011; Kochetov, 2011; O’Brien, 2012), and labial-velar segments show a tendency to change into labials rather than velars (Cahill, 1999; Ohala & Lorentz, 1977). However, despite a large body of evidence demonstrating the integration of vision and other non-auditory modalities in speech perception, only a handful of studies have directly considered this hypothesis (Havenhill & Do, 2018; Johnson, 2015; Johnson et al., 2007; King & Chitoran, 2022; McGuire & Babel, 2012). McGuire and Babel (2012) argue that asymmetries in the misperception of English /θ/ and /f/ may be partly attributable to differences in their visual perceptibility. Whereas /f/ is consistently articulated with visible labiodental constriction, /θ/ may be produced with either a visible interdental constriction or a less-visible dental constriction. They find that /f/ and /θ/ are more reliably discriminated in audiovisual vs. audio-only or video-only conditions, while listener sensitivity in video-only conditions is greater for speakers who produce /θ/ interdentally than for those who produce /θ/ with a non-visible tongue gesture. King and Chitoran (2022) propose that the recent rise in the use of labiodental variants of English /r/ (i.e., [ʋ]) is facilitated by its visually distinct lip configuration. They show that [ʋ] and [w], which have distinct labial configurations (King & Ferragne, 2020), are distinguished equally well in audiovisual and visual-only perception, suggesting that visual cues alone are sufficient for preserving the /r/-/w/ contrast. As labiodental [ʋ] is acoustically similar to /w/ (more so than lingual [ɹ]), they argue that visual cues reinforce the /r/-/w/ contrast, in support of the hypothesis that phonological systems are optimized for audiovisual perceptibility.

When sounds have the potential to be produced with varying combinations of visible and non-visible articulatory gestures, audiovisual perception may also influence patterns of within-speaker variability. In clear speaking styles, for example, speakers make a number of modifications to their speech in an apparent attempt to improve intelligibility for the listener. Clear speech involves a number of global changes (applying to all segments), including increased intensity, slower speaking rate, and larger pitch ranges, among others (for reviews see Smiljanić & Bradlow, 2009; Uchanski, 2005). Segment-specific changes are also observed, e.g., higher-amplitude and higher-frequency spectral peaks for stop bursts and sibilants, longer VOT, and expansion of the F1×F2 vowel space (Bradlow et al., 2003; Cho et al., 2011; Ferguson & Kewley-Port, 2002; Krause & Braida, 2004; Moon & Lindblom, 1994; Picheny et al., 1986). It is not fully resolved to what extent hyperarticulation targets specific phonological contrasts (Buz et al., 2016; Nelson & Wedel, 2017; Schertz, 2013; Tupper et al., 2021; Wedel et al., 2018) as opposed to being the byproduct of greater articulatory effort in general (Ohala, 1994; Wright, 2004). For instance, Bradlow (2002) and Smiljanić and Bradlow (2005) show that similar vowel space expansion occurs in languages with less crowded vowel spaces (Spanish, Croatian) as in those with more crowded vowel spaces (English). While the majority of research on clear speech has focused on acoustic properties, articulatory studies indicate that speakers may increase the velocity, magnitude, and duration of articulatory gestures (Matthies et al., 2001; Perkell et al., 2002). Such adjustments vary on an interspeaker basis, however—Perkell et al. show that some speakers rely on changes to duration or intensity rather than articulatory effort.

Clear speech is usually interpreted as listener-oriented enhancement, often under the framework of H&H Theory (Hyper- and hypo-articulation; Lindblom, 1990). H&H Theory proposes that speakers optimize their speech for perceptibility when they anticipate that the listener’s perceptual needs demand it or when their message is contextually less predictable. Perceptual enhancement competes with speaker-oriented goals such as minimizing articulatory effort, so articulatory gestures may be hypoarticulated when not necessary to convey phonological contrasts. The perceptual advantages of audiovisually-transmitted speech are well established (e.g., Anderson et al., 1997; Gagné et al., 1994, 1995, 2002; Macleod & Summerfield, 1987; Sumby & Pollack, 1954), so some articulatory adjustments may occur, at least in part, to increase their visibility. Under especially noisy conditions, speakers might even optimize their speech purely for visual perceptibility, e.g., if they believe that the auditory/acoustic signal has no chance of recovery by the listener. Previous work suggests that speaking strategies are indeed modulated by vision. Anderson et al. found that speakers’ utterances were auditorily less intelligible when they were aware that their interlocutor was able to see them. Ménard et al. (2016) tested the clear speech articulation of sighted and congenitally blind speakers. They found that while sighted speakers hyperarticulate lip gestures in clear speech, blind speakers were more likely to hyperarticulate lingual gestures. These results suggest that in order to enhance intelligibility, sighted speakers consider how their speech will be conveyed both optically and acoustically.

At the same time, hyperarticulation is not restricted to clear speech and is not necessarily driven by real-time communicative demands. An adjacent line of research has considered how phonetic variability is mediated by the lexicon (Baese-Berk & Goldrick, 2009; Fricke et al., 2016; Munson & Solomon, 2004; Wright, 2004). In ‘contrastive hyperarticulation,’ lexical items in dense phonological neighborhoods or that have minimal pair competitors, are produced with more extreme phonetic features along phonologically contrastive dimensions (Nelson & Wedel, 2017; Schertz, 2013; Wedel et al., 2018). For instance, Baese-Berk and Goldrick (2009) show that words with minimal pair neighbors (e.g., cod and god) exhibit a longer VOT compared to words lacking such a neighbor (e.g., cop vs. *[ɡɑp]). Both perception-oriented and production-oriented accounts for this phenomenon have been proposed. On the one hand, lexical competition may promote hyperarticulation for the purposes of perceptual optimization (Scarborough, 2010; Wright, 2004), as less frequent words and those in denser phonological neighborhoods are more difficult to identify (Luce & Pisoni, 1998). While production-based accounts do not fully dismiss this possibility (Lee & Baese-Berk, 2020), hyperarticulation has also been shown to occur even when potentially competitive lexical items are not relevant to the discourse (Baese-Berk & Goldrick, 2009; N. P. Fox et al., 2015; Wedel & Fatkullin, 2017; Wedel et al., 2018). The finding that hyperarticulation occurs when the speech signal is not likely to be misapprehended calls into question the extent to which it is driven by the listener’s perceptual requirements (or the speaker’s estimation thereof). On this basis, Baese-Berk and Goldrick (2009) argue that lexically-conditioned variability is more fully explained by speaker-internal mechanisms of production. Specifically, activation of the production target induces increased (co-)activation of its phonological neighbors. Successful production requires stronger activation of the target than of its competitors, which results in hyperarticulation. Discourse status strengthens this effect—both the target and its competitors are more highly activated when one is contextually relevant, so hyperarticulation is more extreme.

Most work in this area has focused on the acoustic speech signal, often with unidimensional variables like VOT or the Euclidean distance between vowels (e.g., Bradlow, 1996; Clopper & Tamati, 2014; N. P. Fox et al., 2015; Gahl et al., 2012). Some such measures can readily infer increases to articulatory effort; temporal expansion of VOT or coarticulation (Zellou & Chitoran, 2023) plainly correspond to gestures sustained for a longer period. Yet many other acoustic measures, including a vowel’s position within the F1×F2 space, reflect the combined actions of multiple gestures. To take nasal vowels again as an example, hyperarticulation may not only involve adjustments to the duration or magnitude of velar gestures, but also to tongue height and laryngeal setting, all of which influence the perceived quality of F1. For backness/rounding contrasts, conveyed by differences in F2, hyperarticulation may involve changes to the backness of the tongue, protrusion of the lips, and height of the larynx (Lindblom & Sundberg, 1971; Riordan, 1977). While acoustic distance from one vowel to its competitor is a useful proxy for auditory perceptibility, direct observation of speakers’ articulatory strategies is necessary to obtain additional insight not only into how speakers maintain and enhance phonological contrasts, but also as to whether contrast is influenced by non-auditory perception.

This study examines variability in the strategies used to articulate (and hyperarticulate) vowel contrasts, considering the potential influence of differences in their visibility. Specifically, for vowels distinguished by both backness and rounding, are labial gestures less variable, less likely to be reduced, or more likely to be hyperarticulated than lingual gestures? Three experiments examine inter- and intraspeaker variability of the cot-caught contrast among speakers from Chicago, considering production in clear and normal speech as well as audiovisual perception. cot and caught (/ɑ, ɔ/, henceforth lot and thought [Wells 1982]) are highly variable and have experienced a range of chain shifts and mergers in different regions throughout the past century. The phonological status of their contrast is marginal in most parts of North America, which makes their articulatory characteristics of theoretical interest, particularly in regions where the contrast has not (yet) collapsed.

Until recently, the Chicago variety of English has been characterized by the Northern Cities Shift (NCS), a vowel chain shift that developed during the late nineteenth and mid twentieth centuries in the Great Lakes region (the ‘Inland North’) of the United States (Labov et al., 2006). In its earliest stages, the NCS describes the coordinated movement of the vowels trap (/æ/), lot, and thought, as shown in Figure 1a. Labov et al. (2006) propose that the shift began with the raising of trap, which can exhibit an F1 as low or lower than kit (/i/, Labov 1994). Under their proposal, the raising of trap creates an opening in the vowel space which lot moves forward to fill. They find that Northern Cities-shifted speakers typically exhibit a mean F2 for lot of greater than 1450 Hz, far higher than Peterson and Barney’s (1952) finding of a mean F2 for lot of 1220 Hz for women and 1090 Hz for men. Thomas (2001) and McCarthy (2010) argue that the shift began with the fronting of lot, for which McCarthy (2010) finds evidence in recordings of Chicago speakers born in the 1890s. In either scenario, the movement of trap and lot is followed by the lowering and fronting of thought, which adopts the former position of lot.

These changes have clear motivation under models of acoustic vowel dispersion: Movement of thought occurs in order to fill an open region of the vowel space caused by the fronting of lot. However, because an increase in F2 can be the result of any gesture that shortens the front cavity of the vocal tract, acoustic dispersion alone cannot predict the articulatory strategies used to distinguish vowels along this dimension. Both lip rounding and tongue position influence F2 and have the potential to make equivalent changes to the acoustic output. Majors and Gordon (2008) investigated these alternatives through an analysis of video recordings of two speakers from St. Louis, where the NCS is in effect to some extent. They find that thought is fronted while retaining its rounding, suggesting that thought-fronting and lowering can be accomplished through a repositioning of the tongue alone. However, their analysis did not involve direct observation of tongue position. Using lip video and ultrasound tongue imaging, Havenhill and Do (2018) find that three predicted patterns occur among Metro Detroit speakers. While some speakers produce fronted thought such that it is distinct from lot in both tongue position and lip rounding, others contrast thought from lot through either tongue position or lip rounding alone. These strategies are not equal with respect to their effects on the acoustic output, however. For speakers who distinguish thought from lot by only one articulatory gesture, the vowels are acoustically less distinct than for speakers who contrast them by multiple gestures. Because both single-articulator strategies yield a similar degree of acoustic overlap between lot and thought, Havenhill and Do (2018) also tested their visual perceptibility. They found that the absence of visual lip rounding cues for thought increases the likelihood that listeners will perceive an auditory thought stimulus as lot. If visible rounding cues are important to the identification of lot and thought, then it may follow that articulatory strategies which would eliminate lip rounding will be dispreferred on audiovisual grounds, even if lingual contrast alone is sufficient for differentiating the vowels in acoustics. If so, speakers may then be less likely to reduce the magnitude of labial gestures than to reduce the magnitude of lingual gestures (in normal or hypoarticulated speech), while they may be more likely to increase the degree of lip rounding in hyperarticulated speech.

It is not necessarily the case that speakers in traditionally Northern Cities-shifted regions like Chicago will maintain the lot-thought contrast. Merger of the low back vowels is increasingly common throughout North America and is proposed to be the catalyst for the Low-Back-Merger Shift (LBMS; Becker, 2019; Labov et al., 2016; Roeder & Gardner, 2013). This supraregional chain shift, which has also been referred to as the Canadian/California Vowel Shift (Clarke et al., 1995; Eckert, 2008), Elsewhere Shift (Stanley, 2020), and Third Dialect Shift (Clarke et al., 1995; Labov, 1991), is illustrated in Figure 1b. The shift is proposed to begin with retraction of lot and its merger with thought, which preserves thought as the surviving category (Becker, 2019; Gardner & Roeder, 2022; Labov, 2019). The loss of lot from the acoustic low vowel space (Thomas, 2019) and/or short vowel system (Labov, 2019) motivates the subsequent retraction and lowering of trap, dress (/ɛ/), and kit.

While the LBMS has become widespread throughout much of North America, it remains an open question to what extent it may take hold in regions where the pre-existing vowel systems are less conducive to lot-thought merger (D’Onofrio & Benheim, 2020; Nesbitt & Stanford, 2021; Nesbitt et al., 2019). This is true not only for Chicago, but also the South and the Mid-Atlantic, which have likewise resisted the low back merger but have different phonetic implementations of the contrast. Labov (2019) notes that while the NCS distances lot from thought via lot-fronting, speakers in New York City and the South instead differentiate thought from lot by raising or diphthongizing it, respectively (Kendall & Fridland, 2017; Labov, 1966; Labov et al., 2006; Nycz, 2018; Stanley et al., 2021; Thomas, 2001). He proposes that North American English varieties can be taxonomically organized according to whether the vowel class lot is round or unround, which relates in part to its prior merger with palm in most places. Despite their many differences, the regions just noted are unified by an unround lot. They stand apart from the rest of North America, where roundness of lot encourages its merger into thought. The phonetic quality of the merged category has been described either as [ɑ]-like or [ɔ]-like, perhaps varying regionally (D’Onofrio et al., 2016; Hall-Lew, 2013; Kendall & Fridland, 2017; Kennedy & Grama, 2012), although its articulatory characteristics often go undescribed, being examined mostly through acoustics. Under Labov’s proposal, the rounding distinction between lot and thought (or lack thereof) is a key factor in their merger; the articulatory implementation of these vowels is therefore of interest, as it may inform where low back merger is more or less likely to occur.

This study examines the lot-thought contrast through an acoustic, articulatory, and perceptual study of Chicago English. Three experiments examine inter- and intra-speaker phonetic variability in the articulation of lot and thought, as well as the possible influence of visual cues on perceptibility of the contrast. Experiment 1 focuses on interspeaker variation; that is, whether speakers differ from one another in how they combine lingual and labial gestures to produce lot and thought, and to what extent they vary in the strength of the acoustic contrast. Experiment 2 looks at intraspeaker variation, examining how speakers hyperarticulate lot and thought in clear speech, whether hyperarticulation increases the acoustic distance between them, and whether lingual or labial gestures are preferentially hyperarticulated. Experiment 3 tests whether visual rounding cues aid in the identification of lot and thought through an audiovisual perception experiment.

Eighteen participants (4 men, 14 women) participated in the study, which was conducted at Northwestern University in Evanston, Illinois. Demographic information is presented in Table 1. Participants were natives of the Chicago area, having been born and raised in the region at least through age 18. The age range of participants was 20 to 77 years (M = 46.7, SD = 21.2). All participants had self-reported normal hearing and speech as well as normal or corrected-to-normal vision. One additional participant (a 19-year-old woman not listed in Table 1) also took part in the study but was excluded from analysis because she later reported that she had lived well outside the region for over five years during adolescence.

Speakers are first assessed in terms of their participation in the NCS using benchmarks from the ANAE (Labov et al., 2006), of which two are applicable to the present study. AE1 identifies speakers with advanced trap-raising (F1 below 700 Hz), while O2 identifies speakers with advanced lot-fronting (F2 above 1450 Hz). The fronting of goose is also considered given its potential (socioindexical) association with non-NCS vowel systems (Clopper et al., 2019; D’Onofrio & Benheim, 2020). For present purposes, speakers are categorized by whether goose F2 is front of center (U2 > 1595.5 Hz). Figure 3 provides normalized by-speaker formant means with AE1, O2, and U2 indicated for reference. Measurements were sampled at the maximum F2 (trap) or F1 (all other vowels), the inflection points used by the ANAE.

Overall mean F1 for trap is 693 Hz (SD 89). Eleven speakers satisfy the AE1 criterion, with the most extreme raising seen for CHI003 (female, born 1963). As highlighted in Figure 3, her mean trap F1 is 606 Hz (SD 43), lower than the mean F1 of goat for any speaker (min. 611 Hz) and approaching high front fleece. Her lot and thought are likewise consistent with the NCS; lot is fronted well beyond 1450 Hz (mean F2: 1505 Hz) and is clearly distinct from thought (the mean F2 of which is also high: 1368 Hz). In regions where the NCS is undergoing reversal, younger speakers adopt a continuous or nasal trap system (D’Onofrio & Benheim, 2020; Nesbitt, 2023; Wagner et al., 2016) with raising only in pre-nasal contexts. Here, only one trap item had a non-oral coda (dan) so it was excluded from calculation of the mean. As most speakers thus have relatively raised trap in exclusively oral contexts, its position is generally more NCS-like than LBMS-like. CHI001 (female, born 1994) has the lowest and most retracted trap (F1: 783 Hz, F2: 1706 Hz) and will be seen to have one of the least distinct low back contrasts.

The overall mean F2 for lot is 1432 Hz (SD 141) with the individual means for seven speakers exceeding O2. In contrast to lot, the F2 of which is highly variable, individual F2 means for thought are more tightly clustered around the group mean (1316 Hz, SD 109). The F2 of lot occupies an approximate range of 1250–1650 Hz, nearly encompassing the F2 range of thought (1200–1450 Hz). The speaker with the least distinct lot-thought contrast (in general) is CHI009 (male, born 1998), for whom the mean F2 for lot and thought are 1271 Hz (SD 113) and 1220 Hz (SD 112), respectively. As indicated in Figure 3, he is also among the speakers with the most advanced goose-fronting (F2: 1677 Hz, SD 325), while his trap falls somewhat short of AE1 (F1: 720 Hz, SD 58).

Consistent with variability in the acoustic strength of the lot-thought contrast, speakers also vary as to whether they distinguish the vowels by tongue position. Mid-sagittal tongue contours for lot and thought at 35% of the vowel’s duration are presented in Figure 9 for four speakers. Contours represent midsagittal tongue shape with the tongue front at right. The 95% confidence interval for each spline is shaded; overlap suggests that tongue positions for lot and thought do not differ significantly at that region. CHI003 and CHI010, whose Pillai scores exceed 0.75, produce lot and thought with distinct tongue positions in line with their acoustics. For these speakers, the tongue body for lot is both higher and less retracted than for thought, consistent with earlier descriptions of lot-fronting as also involving raising (Labov et al., 2006, p. 17). Lingual lot-thought contrast is smaller for CHI001, whose acoustic contrast is moderate (Pillai = 0.508), although just-significant trends in the expected direction are apparent. For CHI009, whose Pillai score is indicative of complete merger, the two vowels have identical tongue positions.

Lingual lot-thought differences throughout the full vowel interval are shown as three-dimensional difference surfaces in Figure 10, obtained with itsadug::get_difference(). Like CHI003 and CHI010, the other six speakers with Pillai scores above 0.7 produce lot and thought with clear differences in tongue position. The difference is greatest during the initial 50% interval, coinciding with the typical F2 minimum for thought, at which point acoustic distance between lot and thought is most extreme. The lingual lot-thought difference is not significant at the offset, consistent with the ingliding trajectory for thought F2. As most of these speakers (except CHI005 and CHI017) also produce lot with an F2 exceeding 1450 Hz, lot-thought F2 differences correspond well with tongue position. Speakers with Pillai scores between 0.3–0.7 (CHI004–CHI016) less consistently distinguish lot and thought by tongue position. Within this group, CHI004 and CHI018 are similar to the speakers with high Pillai scores, while CHI001, CHI011, and CHI014 produce thought with greater retraction than lot, but without a significant difference at the tongue front. A marginal difference for CHI012 runs counter to the expected direction; neither CHI012 or CHI016 use tongue position to achieve their relatively weaker acoustic contrasts. Likewise, CHI009 shows no difference in tongue position at any point in the vowel’s duration, consistent with complete merger. On the other hand, CHI013 produces the two vowels with distinct tongue positions throughout their entire duration, despite their acoustic proximity. CHI019 shows some difference in tongue body retraction toward the vowel offset, whereby thought remains retracted while lot begins to front, consistent with her F2 trajectory for thought.

Lingual differences were formally tested with GAMMs fit to two quantitative measures of tongue position. The radial axes of the maximal lot-thought difference were identified from the GAMM difference surfaces, along both the anterodorsal (between 0 and 0.3 radians) and posterodorsal (between –0.7 and –0.3 radians) regions of the tongue. These axes are indicated with dashed lines in Figures 9 and 10. The distance from the probe origin to the tongue surface along each axis was then determined throughout the entire duration of each vowel. Similar to the approach used by Mielke et al. (2017), anterodorsal distance was subtracted from the posterodorsal distance and the result was divided by two, providing a measure of overall tongue body retraction. Larger values for this measure correspond to greater retraction of the tongue body. The y-coordinate for the DeepLabCut-tracked point corresponding to the tongue dorsum (Dorsum1) was used as a measure of overall tongue dorsum height. Higher values indicate a more raised tongue position. These measurements were calculated for all vowels (including /i, æ, u, o/) and z-score normalized. The range of the z-scores therefore also reflect the positions of lot and thought within the articulatory vowel space.

Model summaries for both lingual measures are provided in Table 5. These models confirm that most speakers with Pillai scores above 0.7 produce lot with a significantly less retracted tongue body and a significantly higher tongue dorsum. Visualization of the predicted trajectories for each measure in Figure 11 show that the significant non-linear differences between the two vowels are such that the difference in tongue height and retraction is greatest within the 25–50% interval of the vowel’s duration, during which time thought is the backest and lowest vowel in the entire system. Although visual comparisons of the difference surfaces for speakers with Pillai scores below 0.7 were suggestive of brief or marginal lingual differences, formal testing of the tongue body retraction and tongue dorsum height measures show that lot and thought do not have different tongue positions for most speakers with moderate or low Pillai scores. Notably, however, both CHI013 and CHI019 produce thought with a significantly more retracted tongue body, which is consistent with significant intercept and non-linear differences observed for F2. For CHI019, the difference in tongue body retraction is largest at the offset, which again coincides with the timing of her greatest acoustic difference between the vowels.

While a lingual distinction between lot and thought is consistently observed only for speakers with the highest Pillai scores and most fronted lot, nearly all speakers distinguish lot and thought through differences in lip rounding. Summaries of the GAMM estimates and difference smooths are given in Table 6. Whereas differences in lingual retraction, raising, or both were non-significant for most speakers with Pillai scores below 0.7, all speakers (except CHI009) show significant differences between lot and thought in terms of lip protrusion. In most cases, lot and thought differ in both the overall degree of protrusion and in the non-linear lip trajectories. Figure 12 provides the predicted GAMM smooths for lower lip protrusion with 95% confidence intervals. For three speakers, the non-linear smooth for lot does not significantly differ from thought, while for CHI014 and CHI016, the non-linear smooths have an edf close to 1. For these speakers, lot and thought are distinguished by protrusion, but both the degree of protrusion for thought and the degree of lip spreading for lot are less extreme. Figure 13 shows lip protrusion trajectories from GAMMs fit to all vowels for six speakers. Speakers in the top row, all of whom have clear lot-thought rounding contrasts, produce thought with a maximum degree of protrusion comparable to goose and goat, but with a distinct trajectory at the offset, while lot shows comparable spreading to fleece and trap. For speakers in the bottom row, the trajectory for thought is variable. For CHI012 and CHI016, its shape is more similar to the unround vowels than to the other round vowels, even though thought is significantly more protruded throughout its entire duration. For CHI016, lot is also less spread than trap, whereas the low unround vowels are comparable for the other speakers. Thus, while the magnitude of rounding/spreading gestures for lot and thought may be reduced, rounding alone is responsible for maintaining a modest acoustic contrast, given that neither speaker distinguishes lot and thought through tongue position.

Differences in lip aperture are also variable. For the majority of speakers, lot has a significantly wider lip aperture than thought, although the intercept difference is consistently smaller than that for lip protrusion. Lip aperture trajectories also tend to be more similar, and the difference smooths for several speakers are not significant. Both lot and thought tend to have a larger aperture toward the vowel offset. For CHI008 and CHI006, the lip aperture for thought is larger than for lot, which is attributable to greater outrounding for thought that increases the vertical distance between the upper and lower lips.

This experiment has shown that Chicagoans vary in the strength of the lot-thought contrast, the articulatory gestures used to distinguish them, and in their use of NCS-like vowel targets more generally. Typical NCS vowel systems were observed for most older speakers, who exhibit raising of trap, fronting of lot, and a robust lot-thought contrast. A strong acoustic distinction between lot and thought is found to be associated with differences in both tongue position and lip rounding, as was also observed in Michigan by Havenhill and Do (2018). For older speakers, lot has a fronted tongue position and is unround, while thought has a lower, more retracted tongue position in addition to lip protrusion comparable to goose and goat. Younger speakers, by contrast, are less likely to show an acoustically strong lot-thought contrast and are more likely to show non-NCS features, including a lower trap and more fronted goose. Acoustically-weak contrasts between lot and thought are associated with a less-fronted lot, which is distinguished from thought predominantly (or entirely) through lip rounding. Only one speaker was found to collapse the lot-thought contrast completely, in both acoustics and articulation. For the remaining speakers, the lingual difference between the vowels is either marginal or not significant. While lip protrusion for thought may be somewhat reduced in comparison to the other round vowels, the lot-thought rounding contrast is maintained to some extent for all but one speaker.

Experiment 2 was conducted in the same sitting as Experiment 1, allowing articulatory data for both experiments to be directly compared. All participants in Experiment 1 also participated in Experiment 2, except CHI009, who was unable to participate due to time constraints. The wordlist was a subset of that used in Experiment 1, only including the target vowels lot and thought. Participants were instructed to “speak clearly and with as much emphasis as possible, as though you are correcting someone who misheard you” (cf. Bradlow, 2002; Picheny et al., 1986). Two words containing each target vowel and two containing the contrasting vowel were embedded in the carrier phrase “I said target_x and target_y, not contrast_a and contrast_b.” The items target_x and contrast_a were a minimal or near-minimal pair, as were target_y and contrast_b. For example, with the word pairs nod and sod vs. gnawed and sawed, the participant would say: “I said nod and sod, not gnawed and sawed.” The full list of phrases is provided in supplementary materials. To avoid prosodic differences, only target_x and contrast_a are analyzed. Items alternated such that each word appeared in all four positions, for a total of 36 phrases.

Data processing and analysis followed the same procedures as those described for Experiment 1. However, because Pillai scores are sensitive to sample size (Stanley & Sneller, 2023), scores for the normal speech context were re-calculated after dropping the third repetition of each word. Two tokens of each word for both tasks yielded 36 repetitions per vowel per task (144 tokens in total). As a result, the normal speech Pillai scores differ somewhat between experiments 1 and 2. The reduced token count yields slightly higher scores for most participants, although interspeaker differences in score are generally preserved. Because the aim of Experiment 2 is to examine acoustic and articulatory contrast in comparison with Experiment 1, this decision was made in order to more accurately estimate the difference in acoustic contrast between the two speaking styles.

Dynamic articulatory and acoustic trajectories were analyzed with an additional set of GAMMs. Models from Experiment 1 were expanded with the inclusion of two terms for clear speech in addition to the main term for vowel class, described in more detail below. Rather than by-speaker models, models were fit for three groups of speakers categorized according to their Pillai score in Experiment 1. Speakers with a score above 0.7 were included in the high score group, speakers with a score below 0.3 were included in the low score group, and other speakers were included in the medium score group. As such, models also included random reference-difference smooths for each term by speaker. A set of by-speaker models were also fit in order to visualize formant trajectories and articulatory changes for individual speakers. For visualization of tongue contours, by-speaker models were fit with the same structure as in Experiment 1, where independent tensor product smooths were fit for each level of a combined factor for vowel×task. In addition to the same measures of tongue position and lip rounding used in Experiment 1, Experiment 2 also includes analysis of jaw height in clear vs. normal speech, given that it has potential to serve as an additional visual cue to vowel quality and also to take on socioindexical meaning (Pratt & D’Onofrio, 2017). As previously noted, the position of the short tendon, which attaches the genioglossus to the mental spine of the mandible, was tracked by the DeepLabCut model. Following orientation of the ultrasound images with respect to the occlusal plane, the vertical coordinate of this point was z-score normalized across all vowel tokens and used as a measure of jaw height. Visualization of the short tendon’s position in all vowel contexts is provided for reference in supplementary materials.

Acoustic and articulatory differences for lot vs thought in clear and normal speech are summarized in Figure 22. These figures show each speaker’s predicted maximum acoustic difference (Euclidean distance) between lot and thought in each task, which are compared to articulatory distances calculated in the same fashion, e.g., protrusion2+aperture2 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ \sqrt {{\rm{protrusio}}{{\rm{n}}^2} + {\rm{apertur}}{{\rm{e}}^{\rm{2}}}} \] \end{document} . For all speakers other than CHI011, lot and thought show increased differentiation by lip rounding in clear speech, although the difference is negligible for CHI019. In general, this increase in rounding corresponds to a comparable increase in acoustic distance between lot and thought. While CHI011 and CHI019 show some increase in the lingual lot-thought distinction in clear speech, most other speakers show a decrease in lingual distance.

Figure 24a shows results for the control stimuli, which were visually (in)congruous in terms of vowel height. That is, auditory fleece was presented either with the original (visually congruous) video or with visually incongruous video from its mid counterpart face, while auditory goose was shown with either goose or goat video. Auditory face and goat were likewise shown with either visually congruous mid or visually incongruous high (fleece, goose) video. For the high vowels, fleece and goose, participants identified the stimulus as high in 97.1% of trials. A two-sample t-test for each vowel shows that there is no significant difference between visually high and visually mid stimuli for either fleece (t(23.961) = 0.398, p = 0.694) or goose (t(30.494) = 0.147, p = 0.884). face was identified as high in 5.8% of trials, which is higher than expected. These responses come mostly from a single participant, CHI015, who identified auditorily mid/visually high tokens of face as fleece in 37% of trials. However, as with the high vowels, the effect of visual incongruity was not significant (t(23.299) = 0.77, p = 0.449). Finally, goat was identified as high in 2.8% of trials, with no significant difference between congruous and incongruous stimuli (t(29.97) = –0.599, p = 0.554). In sum, audiovisual incongruity did not have a significant effect on participants’ perception of vowel height.

A significant effect of visual rounding cues was observed, however, for identification of the target lot-thought contrast. Figure 24b presents results for participants who produced a medium or high acoustic lot-thought contrast in Experiment 1. For these participants, auditory lot was correctly identified as lot in 74.2% of trials when presented with visually unround lips. Perception of auditory thought was substantially less accurate; participants in this group identified auditory thought as lot in 45.2% of trials when presented with congruous lip rounding, such that accuracy was just better than chance. When paired with incongruous video of unround lips, auditory thought was much more likely to be perceived as lot (59.8% of trials). The effect of incongruity for lot is comparatively weak: 67.6% of auditory lot stimuli were perceived as lot even when presented with visibly round lips.

For participants who produced an acoustically weak contrast between lot and thought in Experiment 1, perception results are presented in Figure 25. For CHI009, who exhibits complete merger of lot and thought, perception of lot is at chance: 52.5% when presented with congruous unround lips and 50.0% when presented with incongruous lip rounding. CHI009’s perception of thought was also close to chance, but showed a small effect of visual incongruity in the direction opposite to what was predicted. CHI009 heard auditory thought stimuli as lot in 47.5% in trials with congruous rounding, but in only 40.0% of trials with incongruous unrounding. This result suggests that this participant is unable to reliably distinguish between the two vowels in perception, and that his ability to perceive the lot-thought contrast is not aided by visible lip gestures. While CHI019 was also close to chance in her perception of both vowels, she showed the predicted effect of visual cues. When paired with video of unround lot, she responded with lot in 65.8% of auditory lot trials and 57.9% of auditory thought trials. This result is consistent with this speaker’s production of a significant rounding distinction between lot and thought in Experiment 1, even though her acoustic contrast was weak and she only marginally increased the lip rounding contrast in Experiment 2.

A mixed effects logistic regression model (Table 11) was run for all participants who contrast thought from lot with lip rounding in their own speech production, i.e., all participants except CHI009. Fixed effects included auditory vowel class, visual congruity, and their interaction. By-participant random slopes for vowel class and visual congruity were included, as were random intercepts for talker and item. A significant negative coefficient for auditory vowel class indicates that stimuli containing auditory thought are less likely to be perceived as lot than stimuli containing auditory lot. The non-significant main effect of incongruity confirms that lot is not perceived as thought when produced with visually round lips. A significant interaction of vowel class and visual congruity shows, however, that thought is more likely to be perceived as lot when produced with unround lips.

Figure 26 shows response times by vowel class and visual congruity. Response times for lot and thought were significantly longer than for any other vowel. A linear mixed effects model (Table 12) fit to log response time, with random intercepts for participant, talker, and item, shows that vowel class significantly predicts response time (χ²(5), p < 0.001). Pairwise post hoc comparisons show that response times for each of lot and thought is significantly longer than for all non-low vowels (p < 0.001), but that lot and thought do not significantly differ from each other (p = 0.9747), nor do any of the non-low vowels differ from one another. Inclusion of visual congruity as a predictor does not improve model fit (χ²(1), p = 0.628), indicating that response time does not differ for stimuli presented with video that is incongruous in terms of rounding (lot, thought) or height (all other vowels).

Experiment 3 confirms that visual rounding cues influence listener identification of lot and thought. Stimuli containing auditory thought were significantly more likely to be perceived as lot when presented with unround lips. As thought in many cases exhibits significant acoustic overlap with lot, visual perception of lip rounding may be one of the primary means by which listeners distinguish lot and thought. If so, perceptual benefits conferred by lip rounding may contribute to a bias toward visually distinct articulatory strategies and allow for greater spectral overlap between lot and thought without complete collapse of the contrast. In general, listeners were found to exhibit a bias toward perceiving lot when presented with auditory thought stimuli, regardless of the visual stimulus. A nearly identical bias toward lot was also seen by Havenhill and Do (2018), who tested participants on different recordings of the same nonce words used here. In that study, participants identified auditory lot stimuli as lot in 65–75% of trials regardless of visual lip rounding cues. Even when thought was presented with congruous rounding, participants identified the vowel as lot in over 35% of trials, increasing to 55% of trials when presented with incongruous unrounding. A contributing factor for this bias is likely frequency: Mines et al. (1978) show that the token frequency of lot in conversational American English is approximately twice that of thought. As the words presented to participants were nonce words, a frequency-driven bias toward lot is likely stronger than what would be observed for identification of real thought words. It may also be the case that for auditory lot tokens with a relatively high F2, visible rounding is insufficient to override the auditorily perceived frontness of the vowel. For thought, which is acoustically more back, the auditory cues are inherently more ambiguous. The results of Experiment 1 show that thought rarely, if ever, exhibits high F2 values, while lot ranges from a low back to low central vowel. A low vowel with an F2 around 1200–1300 Hz may thus correspond either to thought or to a back token of lot; in that case, visual cues (and in real speech, information obtained from discourse and syntactic context) may determine what listeners perceive. As this experiment tested only audiovisual perception and did not include audio-only or visual-only tasks, it may also be that case that listeners will show greater reliance on visual cues under noisier conditions, including for lot. Future work may further explore the trading relationship between auditory and visual perception by varying the amount of auditory noise, visibility of the lips, and acoustic quality of the vowel.