2.2. Results

2.2.1. Aggregate results

A1-P0 values were analyzed with a mixed-effects linear regression model using the lmer() function in the lme4 package (Bates et al., 2015) in the R program for statistical computing (R core team). Estimates for degrees of freedom, t-statistics, and p-values were computed using Satterthwaite approximation with the lmerTest package (Kuznetsova et al., 2017). The model was run on centered and scaled A1-P0 values and included fixed effects of Vowel Type (three levels: oral [reference level], coarticulated, nasal), and Timepoint (continuous, coded as 1, 2, 3; centered), and the two-way interaction between the two effects. The random effects structure of the model included by-speaker and by-word random intercepts and by-speaker random slopes for Vowel Type did not lead to convergence. All categorical variables were treatment coded. (lmer syntax: A1-P0~Vowel Type*Timepoint + (1|Speaker) + (1|Word)).

Figure 2 provides the average A1-P0 values over normalized time. As expected, nasal vowels contain greater acoustic vowel nasalization (lower A1-P0 values) than vowels overall [coef. = –1.4, t = –24.8, p < 0.001]. Coarticulated vowels also contained greater acoustic nasality (lower A1-P0) than oral vowels [coef. = –0.3, t = –4.9, p < 0.001]. There was also an interaction between Vowel Type and Timepoint, such that nasalization increases in nasal vowels over time [coef. = –0.2, t = –7.9, p < 0.001], and also increases in coarticulated vowels over time [coef. = –0.1, t = –2.5, p < 0.05]. No other effects were significant.

Figure 2

Mean acoustic vowel nasalization (A1-P0 dB) values across three equidistant timepoints for oral vowels (from CVC words), coarticulated vowels (from CVN words), and nasal vowels (from CṼ words).

To assess the relationship between nasality and vowel quality, two separate lmer models were run on log F2 and F1 values taken at the midpoint (centered and scaled). Each model contained fixed effects of Vowel Type (three levels: oral [reference level], coarticulated, nasal), and Phonemic Vowel (three levels: /ɑ/ [reference level], /ɛ/, /ɔ/), and the two-way interaction between the two effects. The random effects structure of both models included by-speaker and by-word random intercepts and by-speaker random slopes for Vowel Type and Vowel. All categorical variables were treatment coded. (lmer syntax: Formant~Vowel Type*Vowel + (1 + Vowel Type + Vowel|Speaker) + (1|Word)).

Figure 3 displays mean log F1 and F2 values. In the model for F2, entirely as expected, there was an effect of Vowel: overall, /ɛ/ has a higher F2 than /ɑ/ [coef. = 0.9, t = –2.3, p < 0.05], consistent with a fronter tongue position. While there was not a main effect of [ɔ] on F2, there was a marginal interaction between Vowel Type and Vowel involving /ɔ/ [coef. = 1.3, t = –2.1, p = 0.05], such that the difference between oral and nasal /ɔ/ is smaller than that between oral and nasal /ɑ/.

Figure 3

Mean and standard errors of log F1 and F2 values (at midpoint) for oral vowels (from CVC words, squares), coarticulated vowels (from CVN words, triangles), and nasal vowels (from CṼ words, circles). Colors reflect oral and nasal counterparts: green = [ɛ] and [æ̃], blue = [ɔ] and [õ], red = [ɑ] and [ɔ̃].

In addition, the model for F2 revealed a main effect of Vowel Type: nasal vowels contain lower F2 values than oral vowels [coef. = –1.3, t = –3.3, p < 0.01], consistent with a more retracted tongue position. Meanwhile, oral and coarticulated vowels do not differ significantly [p = 0.9]. No other main effects or interactions were significant.

In the model for F1, again as expected, there was a main effect of Vowel such that the mid vowels have lower F1 values relative to the low vowel (for the front mid vowel: [coef. = –1.2, t = –4.9, p < 0.001]; for the back mid vowel: [coef. = –1.0, t = –4.4, p < 0.001]), consistent with a higher tongue position.

In addition, the F1 model also revealed that nasal vowels contain a lower F1 than oral vowels [coef. = –0.7, t = –3.5, p < 0.01], consistent with a higher tongue position. Yet, this was mediated by an interaction between Vowel Type and Vowel: /ɛ/ has a higher F1 value when it is nasal, compared to when it is oral [coef. = 1.5, t = 4.4, p < 0.001], consistent with a lower tongue position. There were no main effects or interactions involving the comparison between the oral and coarticulated vowels.

Finally, with respect to vowel duration, vowels were longest in CṼ words (139 milliseconds), and shorter in CVC (110 ms) and CVN (113 ms) words. We ran a mixed effects linear regression model on vowel duration (logged, centered, and scaled) with a fixed effect of Vowel Type (three levels: oral [reference level], coarticulated, nasal), including by-speaker and by-item random intercepts and by-speaker random slopes for Vowel Type (lmer syntax: Vowel Duration ~Vowel Type + (1 + Vowel Type|Speaker) + (1|Word)). The model confirmed that nasal vowels were longer than oral vowels [coef. = 0.92, t = 3.9, p < 0.001]. There was no difference in duration between oral and coarticulated vowels [coef. = 0.22, t = 0.5, p =0.6].

2.2.2. By-speaker results

Figure 4 provides the mean A1-P0 values across the thirty speakers. Significant variation is present. Some speakers produce comparable amounts of nasality in CVC versus CVN contexts (i.e., circles and triangles nearly overlap, such as for S009), suggesting that they coarticulate only to a very limited degree, if at all. Meanwhile, other speakers produce similar amounts of nasality in CṼ versus CVN contexts, suggesting that they coarticulate to a greater degree (i.e., triangle and squares nearly overlap, such as for S025).

Figure 4

Mean acoustic vowel nasalization (A1-P0 dB) values (averaged across three vowel timepoints) for each speaker for oral vowels (from CVC words), coarticulated vowels (from CVN words), and nasal vowels (from CṼ words).

2.3 Interim summary

Our production results replicate previous findings from the literature. As expected, acoustic nasality is greatest in CṼ contexts, and relatively diminished in other contexts. While this pattern is broadly consistent with an account in which speakers deliberately limit nasal coarticulation (cf. Manuel, 1990), our findings also show that CVN contexts still exhibit significantly more nasality than CVC contexts, suggesting that some degree of coarticulation does occur. Also as expected, CṼ contexts differ significantly in F2 and F1, compared to CVC and CVN contexts. This finding is consistent with previous accounts suggesting that French oral and nasal vowels exhibit differences not just in nasality, but also in quality. We also observe individual variation in coarticulatory patterns.

3.2. Results

3.2.3. Individual differences across speakers

Figure 6 provides mean accuracy rates for the thirty individual speakers who provided stimuli, ordered by mean identification rates for CVN contexts.

Figure 6

Proportion correct identification by stimulus type for each individual speaker.

For phonemically nasal vowels from CṼ contexts, accuracy is very high and does not vary much across speakers, similar to what we see in the aggregate results. So, essentially all speakers produce nasal vowels that are unambiguous, in the sense that they are not confusable with either oral or coarticulated vowels.

For the oral vowels from CVC contexts, performance is overall lower, again similar to what was seen in the aggregate pattern. However, the range of values across speakers, from 60% to 90% accuracy, is larger than for the nasal vowels, suggesting that there is more variation in how speakers provide cues to orality.

Finally, for the coarticulated vowels from CVN contexts, we see the largest range of variation across speakers. In the lower accuracy ranges depicted on the left side of Figure 6, stimuli from CVN contexts are correctly identified as CVN lexical items around 35–40% of the time. Technically these results are at chance, since listeners choose from three options. However, the confusion matrix in Table 2 makes clear that the nasal vowel is not really in competition here. Instead, for stimuli from these speakers, there is greater competition from the CVC category. Moving toward the middle of Figure 6, there are also many speakers whose stimuli from CVN contexts are identified in the 50% accuracy range. For these speakers, CVN vowels are truly ambiguous between CVN and CVC. Finally, toward the right side of Figure 6, we see several speakers whose stimuli are accurately identified as originating from CVN words over 60% of the time.

To assess the relationship between acoustic properties of nasal coarticulation and the talker-specific identification patterns, we ran a linear regression model on mean accuracy from CVN contexts only. This revealed a significant effect, depicted in Figure 7 [coef. = –0.01, t = –3.8, p < 0.001].

Figure 7

Correct identifications of coarticulated vowels by speakers’ A1-P0 mean.

Finally, to assess the role of vowel quality in perception, we calculated the difference between mean log F2 values for coarticulated versus nasal vowels, for each individual speaker. The difference between mean log F1 values for coarticulated versus nasal vowels for each speaker was also calculated. We ran two separate linear regressions, with speaker mean accuracy for CVN items as the dependent variation and the speaker-specific formant difference as the predictor variable. The F2 model did not compute a significant effect [coef. = 0.04, t = 0.3, p = 0.8]. There was also not a significant effect in the F1 model [coef. = 0.05, t = 0.5, p = 0.6].

4.1. Interplay between coarticulation and the grammar

We interpret our results as a first step in resolving an important conundrum in the interplay between physical events and abstract categories: namely, while coarticulatory cues are a rich source of information for speech perception, their effects can be modulated by the phonological grammar. As noted in the Introduction, the study by Beddor and colleagues (2013) provides a representative example of how coarticulatory cues help listeners interpret speech. In an experiment with American English stimuli in a visual world paradigm, they showed that listeners fixated more quickly on a visual image corresponding to a target such as spoken bent when coarticulatory nasalization occurred early in the vowel, rather than late. These findings are in concert with a long line of previous studies (starting e.g., with Malécot, 1960), and demonstrate that listeners respond to coarticulatory cues as soon as they become available.

Moreover, other studies have suggested that such conclusions must necessarily be language-specific. For example, in an experiment with CVC and CVN stimuli in a gating paradigm, Lahiri and Marslen-Wilson (1991; see also Kotzor et al., 2022) showed that while English listeners interpreted nasality on the vowel as evidence for the presence of a nasal coda, Bengali listeners interpreted such cues as evidence for a phonemically nasal vowel. On this basis, Lahiri and Marslen-Wilson (1991) argued that when a language does not contain phonemically nasal vowels, as in English, listeners interpret nasality as evidence for coda N; but when a language does contain phonemically nasal vowels, as in Bengali, listeners interpret nasality as evidence for Ṽ.

Our findings introduce an important refinement to this scenario. Like Bengali, French contains phonemically nasal vowels. However, we have shown that French listeners can interpret nasality as evidence for VN. As we speculated earlier, the differences between French versus Bengali listeners most likely stem from an important fact about French, namely that oral and nasal vowels are distinguished not just by nasalization, but also by differences in quality. The same is not true of Bengali, where nasal vowels are equivalent in quality to their oral counterparts (e.g., Ferguson & Chowdhury, 1960). Thus, in mapping the speech signal onto a particular vowel phoneme, Bengali listeners must rely solely on nasality cues, whereas French listeners interpret nasality and quality in tandem (Delvaux, 2009). More broadly, both in the current work and in previous works (e.g., Beddor, Krakow, & Goldstein, 1986; Benguerel & Lafargue, 1981; Carignan, 2014; Delvaux et al., 2008; Delvaux, 2009; Dow, 2020; Krakow et al., 1988; Solé, 1992, 2007), French provides us with an opportunity to reconsider the characteristics of nasality as a phonological feature, and to highlight its multifaceted nature.

Recall that our analysis of perception results for CVN contexts did not show a significant correlation between F2 differences across CVN versus CṼ contexts, on the one hand, and listener accuracy, on the other. In other words, greater differences between oral versus nasal vowel qualities did not necessarily lead to greater accuracy in distinguishing CVN and CṼ stimuli. However, this correlation is not strictly necessary in order for our basic logic to hold: as long as the baseline quality difference between oral versus nasal vowels remains steady – and our production results show that this is indeed the case – listeners should be able to interpret nasal cues against the backdrop of quality cues. Future work could probe this issue further by systematically exploring how variation in vowel quality affects the perception of coarticulation in VN contexts, similar to the manner in which Delvaux (2009) approached CṼ contexts.

A particularly interesting issue, which we were not able to address in the current study, concerns nasal coarticulation on high vowels (e.g., in contexts such as fine [fin] ‘thin, fine’ and lune [lyn] ‘moon’). Previous authors have reported that coarticulation is greater in these contexts, compared to non-high vowel contexts (Delvaux et al., 2008; Dow, 2020), possibly due to the fact that the inventory of French lacks phonemically nasal high vowels (e.g., *[fĩ], *[lỹ]). On this basis, we could hypothesize that, when listening to vowels that contain acoustic nasality, French listeners should make stronger predictions about the presence of N when the vowel is high, compared to when it is non-high.

4.2. Individual differences

In studying the perception of coarticulation, previous researchers have typically created stimuli using cross-splicing and/or acoustic manipulation. While useful, these techniques probe only a narrow range of human behavior. For example, if we use cross-splicing to create matched versus mismatched stimuli (i.e., CV_NASALN versus CV_ORALN), then we probe only how listeners respond to those two types of stimuli. By contrast, a distinctive aspect of the current study is that we presented listeners with stimuli “from the wild” – that is, with stimuli whose patterns of nasal coarticulation varied naturally among thirty different speakers of Metropolitan French. This allowed us to probe how listeners responded to a wide range of stimuli, which more closely resembles the challenges posed by everyday listening. Crucially, this approach also allowed us to show that certain speakers of French have stronger nasality cues in CVN contexts than others, and that listeners make use of these cues when they are present.

This approach closely resembles the methodology employed by Zellou (2022) for American English. In that study, sixty different speakers produced CVC and CVN words (e.g., bed and ben). Their recordings were used in a perception experiment, in which listeners heard CV stimuli with the coda removed, and identified the corresponding word (bed or ben?). Results showed that some speakers’ CVN tokens were (incorrectly) identified as CVC words, while other speakers’ CVN tokens were (correctly) identified as CVN words. American English does not have phonemically nasal vowels, and anticipatory coarticulation in CVN sequences is generally long and strong (e.g., Cohn, 1990). Nevertheless, there is substantial variability from one speaker to the next (Beddor et al., 2018; Styler, 2017; Zellou, 2017; Zellou & Brotherton, 2021). Zellou (2022) showed that this variability exerts direct consequences in listeners’ perceptual choices. That is, American English listeners make use of nasality cues when they are present. (We note that Beddor et al. (2013) and Coetzee et al. (2022) reached similar conclusions, using stimuli that varied categorically between early versus late onset of nasalization.)

Our current pattern of results for Metropolitan French is essentially identical to what Zellou (2022) found for American English: that is, some speakers’ CVN tokens were (incorrectly) identified as CVC words, while other speakers’ CVN tokens were (correctly) identified as CVN words. Unlike American English, however, French does have phonemically nasal vowels, and anticipatory coarticulation in CVN sequences is generally short and weak (Delvaux et al., 2008; Dow, 2020). Despite this rather profound difference, the basic behavior of French listeners is nevertheless equivalent to that of American English listeners. This is a novel finding which suggests the broad hypothesis that, regardless of phonological grammars, listeners always use coarticulatory cues, when they are present, to make perceptual choices. To test this hypothesis in subsequent studies with different languages and participant populations, we emphasize that it will be crucial to present a wide range of naturally-varying stimuli, as we have done here, in order to produce the broadest and most accurate characterization of listener behavior.

4.3. Theoretical implications

Our findings contribute new evidence to certain debates within the speech perception literature. In particular, different frameworks have made differing predictions about how listeners should interpret nasality when it occurs in the signal, particularly for those languages which contain both phonetically and phonemically nasalized vowels, such as French and Bengali. Under a surface-based account, listener behavior simply follows the distributional cues: if nasalization on the vowel is typically followed by a nasal consonant, then listeners will interpret nasalization as a cue to N (for discussion, see Chong & Garellek, 2018). By contrast, under a representational account as laid out in the Featurally-Underspecified Lexicon (FUL) model, listeners will interpret nasalization as a cue to Ṽ (Kotzor et al., 2022; Lahiri & Marslen-Wilson, 1991). This is because the nasalization in the signal is a “match” for CṼ contexts, which underlyingly contain the feature [Nasal]. Meanwhile, nasalization in the signal is a “no-mismatch” for both CVN and CVC contexts, which are underlyingly unspecified and do not contain the feature [Nasal].

Our perceptual results are not entirely consistent with the representational account, at least as it currently stands. Listeners did interpret CṼ stimuli as CṼ words, as predicted by FUL. However, listeners did not interpret CVN stimuli as CṼ words, a finding that would seem to contradict the predictions of FUL. Instead, listeners overwhelmingly interpreted CVN stimuli as either CVC or CVN words. That is, despite the presence of nasality in the signal, they did not map this to a lexical vowel with an underlying [Nasal] feature. To some extent, of course, this result may be due to the fact that the French CVN words contained relatively small amounts of nasalization as well as distinct quality differences across CVN and CṼ vowels, as demonstrated by our production study. Nevertheless, even when the magnitude of nasalization did increase due to natural variation among speaker productions, listeners did not become more likely to select CṼ words. Instead, as our correlation analysis showed, they became more likely to select CVN words. Note, meanwhile, that although our listeners did not map acoustic nasality onto a lexical [Nasal] vowel, they did correctly use nasality as evidence for the presence of a lexical [Nasal] consonant. This suggests that a refined FUL model could potentially make correct predictions for languages like French, particularly if it were to incorporate temporal information, as well as the presence of quality differences between oral and nasal vowels. For instance, if representations for [Nasal] vowels in French included both vowel nasality and vowel quality features, this could explain the lack of confusability by listeners between CVN and CṼ vowels in the current study.

Whether our perceptual results support a surface-based account remains an open question, because the nature of the distribution itself has yet not been fully established. On the one hand, summary trends show that very little nasalization occurs on vowels in CVN contexts, both in our own data and in previous studies (Delvaux et al., 2008; Dow, 2020). On the other hand, some speakers produce greater amounts of nasalization than others do, again in our own data and in previous studies (Styler, 2017). To really pin down the distribution, however, we must ask whether this variation among speakers is principled. For instance, is it consistently affected by sociolinguistic factors such as speaker gender and social class, and/or by stylistic factors such as clear versus casual speech? If the answer to any of these questions is yes, then we would generate an interesting set of predictions. For example, if female speakers of French are generally more likely to exhibit nasal coarticulation than male speakers of French, then a distributional account predicts that listeners will be more likely to make use of coarticulatory cues when listening to a female voice, compared to a male one. Coetzee et al. (2022) tested this hypothesis in a somewhat different context — namely, predictable differences in the amount of nasal coarticulation between two socio-ethnic varieties of Afrikaans — but did not find evidence that listeners adjusted their perceptual strategies based on dialect information alone. The same question could be asked for vowel quality, as well – are there differences across regional varieties in vowel quality across categories that listeners are able to make use of? Our F1 and F2 models confirm that our listeners did not rely on vowel quality (at least not exclusively) to identify the words. But, could listeners still have used the nasality in the signal against the backdrop of vowel quality differences to identify a phonemic nasal vowel? If yes, it would also predict individual perception/production differences whereby some speaker-listeners would rely more heavily on nasal coarticulation, and others more heavily on vowel quality differences, but never on just one of these alone. These issues can be further explored in future work.

4.4. Avenues for future work

Future research on French could employ time-sensitive measurements, such as eye-tracking or gating, in order to probe listeners’ perception of nasality at various time points within the vowel. To our knowledge, only a couple of previous studies have employed these techniques. Desmeules-Trudel and Zamuner (2019) used a visual world paradigm to examine how Canadian French listeners responded to changes in the temporal duration of nasality on vowels. Their results showed that as the nasality duration decreased, listeners were more likely to select a picture depicting a CVN lexical item, compared to a picture of a CṼ lexical item; notably, however, the eye fixation results were not comparable, since proportion of fixations to CVN versus CṼ pictures was not affected. A similar experiment using Metropolitan French, whose vowel system differs in crucial ways from Canadian French, could shed further light on our current results and also allow for a more direct comparison with the findings of Beddor et al. (2013).

Meanwhile, Ingram and Mylne (1994) examined French listeners’ responses to CVC ~ CVN ~ CṼC words, such as tâte ~ tante ~ tanne, using a gating paradigm. Their descriptive results, which were summed over all gates, showed a pattern of results that are different from what we report here. Specifically, they indicated greatest accuracy for CVC, then CṼC, with lowest accuracy for CVN, whereas our own results showed greatest accuracy for CṼC, then CVC, and lowest accuracy for CVN. Ingram and Mylne (1994) do not report any information about their acoustic stimuli (such as who recorded them, or what acoustic characteristics they possessed), and their French listeners resided in Australia, rather than in a French-speaking country. These factors make it difficult to speculate about why our results differ from theirs. Suffice to say that future experiments with gating paradigms could nevertheless shed further light on our current results, and allow for a more direct comparison with the findings of Lahiri and Marslen-Wilson (1991) and Ohala and Ohala (1995), as well as Kotzor et al. (2022).

The use of eye-tracking and/or gating could also help to address certain limitations of the current study. In our perceptual experiment, for example, listeners heard stimuli of the shape CV, since they were presented with recordings in which the original coda, if present, had been excised. This means that while the CVC and CVN stimuli underwent an experimental manipulation, the CṼ stimuli did not, a fact which may have rendered them easier to identify. Note, in addition, that the vowels [ɑ] and [ɔ] rarely occur in open syllables in French (Carignan, 2014), raising the possibility that listeners may have had difficulty interpreting truncated CV sequences such as [fɑ] and [tʀɔ] (from [fɑn] ‘wilt-1-3person’ and [tʀɔɲ] ‘face-Familiar’, respectively). Furthermore, in many languages including French, nasal vowels are typically longer than oral vowels (Delvaux, 2009), and this was also the pattern that we found in our own stimuli, raising the possibility that our participants may have distinguished between VN and Ṽ not only on the basis of vowel quality, as we have speculated, but also on the basis of overall duration.

Another direction for future work is to explore individual differences in the relationship between the production and perception of nasality cues across French speakers. There is some work starting to look at this question. Rodriquez et al. (2023) found that French speakers who use nasalization to a lesser degree to signal Ṽ vs. VN tend to have longer and especially more variable anticipatory nasalization in CVN tokens than those speakers who use nasalization stronger to signal nasal vowels. Thus, this presents a ripe direction for future work.

Finally, future work could also address the question of compensation for coarticulation. Although we have cited previous evidence that listeners use coarticulation as a cue to the presence of particular segment types, there is also ample evidence that listeners “undo” coarticulation in the presence of the segment that caused it (e.g., Mann & Repp, 1980). For example, while listeners are highly likely to perceive a phonemically nasal vowel in the context CṼC, they are less likely to do so in the context CṼN, because they attribute some or all of the vowel’s nasality to the following consonant (e.g., Kawasaki, 1986). Note that in the stimuli for the current study, we excised the coda, which “hid” the source of coarticulatory effects from listeners and rendered it infeasible to assess any potential effects of compensation. Different stimuli and alternative experimental tasks could allow us to assess these effects, an important goal given the evidence that compensation – like the use of coarticulatory cues – may also be modulated by the phonological grammar of a language (Beddor & Krakow, 1999).