The combined effects of contextual predictability and noise on the acoustic realisation of German syllables

: Speakers tend to speak clearly in noisy environments, while they tend to reserve effort by shortening word duration in predictable contexts. It is unclear how these two communicative demands are met. The current study investigates the acoustic realizations of syllables in predictable vs unpredictable contexts across different background noise levels. Thirty-eight German native speakers produced 60 CV syllables in two predictability contexts in three noise conditions (reference ¼ quiet, 0 dB and (cid:2) 10 dB signal-to-noise ratio). Duration, intensity (average and range), F 0 (median), and vowel formants of the target syllables were analysed. The presence of noise yielded signiﬁcantly longer duration, higher average intensity, larger intensity range, and higher F 0. Noise levels affected intensity (average and range) and F 0. Low predictability syllables exhibited longer duration and larger intensity range. However, no interaction was found between noise and predictability. This suggests that noise-related modiﬁcations might be independent of predictability-related changes, with implications for including channel-based and message-based formu-lations in speech production. V C 2022 Author(s). All article content, except where otherwise noted, is licensed under a Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).


I. INTRODUCTION
Linguistic units, e.g., words, syllables or phonemes, occurring in predictable contexts, tend to undergo more reduction than those in less predictable contexts (Aylett and Turk, 2004;Crocker et al., 2016;Frank and Jaeger, 2008) (also known as source coding; see Pate and Goldwater, 2015) but this tendency might be counteracted in noisy environments when it is difficult for listeners to anticipate a message from the degraded signals. In response, speakers often enhance the signals. However, this need contradicts the tendency for speakers to reduce in predictable contexts. This then raises the question as to whether predictability-based modifications will be attenuated or exaggerated in noisy environments. The current study investigates the acoustic realisations of syllables in predictable vs unpredictable contexts across different background noise levels.
Speakers often adapt their speech to different environments, e.g., in a factory or a stadium (see Cooke et al., 2014, for a review). It is generally accepted that speech will be automatically made louder in noisy environments, with the resultant noise-induced speaking style referred to as Lombard speech (Brumm and Zollinger, 2011;Lombard, 1911). This style is considered an adaptation to protect against potential information loss in a non-ideal channel during signal transmission (also known as channel coding; see Pate and Goldwater, 2015). A number of acoustic features have been identified that characterize this speaking style: increased vocal effort, increased intensity, increased fundamental frequency (F0), e.g., Patel and Schell (2008), slow speaking rate, long duration, e.g., Lu (2010), and high first formant (F1) of vowels (see Junqua, 1996, for a review). These characteristics have been reported to help listeners identify speech in challenging communicative settings, suggesting acoustic enhancement of signals (Hazan and Simpson, 2000).
Lombard speech modification does not manifest uniformly across different segment types or linguistic units. Rather, Lombard speech shows up in units that are critical to intelligibility. For instance, speakers emphasize vowels more than consonants in a noisy environment, presumably the former is more critical to speech audibility (Garnier and Henrich, 2014). Content words undergo larger modifications than function words in terms of fundamental frequency, syllable duration (Patel and Schell, 2008), and lexical stress contrastivity (Arciuli et al., 2014). Moreover, speakers differentially increase vocal effort with levels of noise intensity (Ngo et al., 2017;Wakao et al., 1996). The study by Zhao and Jurafsky (2009) examined the effects of word frequency and noise on the acoustic realisation of Cantonese tones. Their results showed an overall increase in F0 for all tones in noise, but they only observed the effect of word frequency for mid tones, with higher F0 in low-frequency than high-frequency words. This is because mid-tones are hard to distinguish perceptually and it will be more difficult to identify an unpredictable (i.e., low frequency) word with mid tones without making it acoustically salient. These previous studies then suggest that speech is modified to serve some communication goals for the benefit of listeners.
Word frequency is one measure for predictability (Ernestus, 2014). Yet predictability goes beyond word frequency, e.g., local contextual predictability. Local contextual predictability is the predictability of a unit, given its preceding (or following) units, referred to as surprisal (Shannon, 1948). Surprisal quantifies information in terms of bits as the inverse of the unit's log probability given the local context: Sðunit i Þ ¼ Àlog 2 Pðunit i jContextÞ (Hale, 2016).
Predictability has been shown to modify speech output at multiple levels. For instance, speakers tend to shorten word duration for predictable, but not for less predictable, messages (Buz and Jaeger, 2016). Such predictability effect is often interpreted under the notion of achieving a speaking intent with minimal effort. Conversely, hard-to-understand units tend to have longer duration, possibly resulting from explicit encoding to improve intelligibility (Gahl et al., 2012;Jaeger, 2010). In addition, American English vowels were found to be more centralized in contextually more predictable syllables (Aylett and Turk, 2006), in line with the idea that vowels are more likely to be weakened or deleted in high-frequency than low-frequency words (Bybee, 2001(Bybee, , 2002. In general, vowels were more dispersed in the vowel space when they were contextually less predictable as in Malisz et al. (2018) and this effect was observed in six different languages (American English, Czech, Finnish, French, German, and Polish) across three speaking rates (slow, normal, and fast).
From the information theory perspective (Shannon, 1948), message formulation (source coding) aims at representing information as accurately as possible in as few bits as possible, while channel coding aims at protecting information from transmission loss over a non-ideal channel. To the best of our knowledge, no study has examined the effects of channel coding (e.g., Lombard speech) and source coding (e.g., local contextual predictability) on the acoustic realisation of a syllable. The primary goal of this study was to examine any combined effects of background noise and contextual predictability in German syllables. We extracted duration, average intensity, intensity range, and F0 features from the syllable, and F1, F2, and F2-F1 difference from the vowel. Throughout this paper, we refer to features extracted from the syllable as syllable-based features and vowel formants as vowel-based features. Those features were chosen as the metrics for speech modification because they are known to correlate with the Lombard effect and some of the metrics have been shown to be sensitive to contextual predictability (e.g., Boril and Poll ak, 2005;Brandt, 2019;Lu and Cooke, 2009).
Our research question was how syllables with different contextual predictability are acoustically realised as a function of noise. We expected syllables with high surprisal or in noise to be articulated with care when compared to their counterparts with low surprisal or without noise, respectively. In addition, if speech is modified to serve some communicative goals for the listener, it is more likely for modification to be made on the least predicted element. Therefore, we also expected acoustic modifications to be larger for syllables with high surprisal than those with low surprisal in noisy conditions.

A. Participants
Thirty-eight native German speakers with no known hearing or speech impairments were recruited (12 M, 26 F; average age ¼ 27.6, 19-60 years) to take part in a reading aloud task. For recruitment, no upper age limit was imposed, in consideration of the observation that older adults with age-typical mild hearing loss do not differ from young adults in how they modify their speech in background noise (Hazan et al., 2018).

B. Stimuli
Twenty stressed CV syllables were created by combining one of the plosives: /p, b, d, k/ and one of the vowels: /a:, e:, i:, o:, u:/. The CV syllables, consisting of two surprisal groups (high vs low), were crossed with three whitenoise conditions (reference ¼ no noise, 0 dB and À10 dB signal-to-noise ratio, SNR), resulting in a total of 60 target syllables. Each target syllable formed part of a polysyllabic word in a sentence context, which was chosen from the DeWaC corpus (Baroni and Kilgarriff, 2006;Brandt et al., 2017). The stimuli were pseudo-randomized for presentation.
A syllable-based language model was used to estimate the probability of a unit given its previous local context, based on large text corpora. The choice to train the language model (LM) with syllable as a unit was motivated by the importance of the syllable as a processing unit in language production and perception (see Krakow, 1999, for a review), and as demonstrated in different phonetic encoding for high frequency vs novel syllables in B€ urki et al. (2015). Although previous studies, e.g., Bell et al. (2009) found an effect of backward predictability on the duration of lexical and function words in American English, a more recent study by Tang and Bennett (2018) found a stronger effect of forward predictability on word duration in Kaqchikel Mayan. Considering the inconsistent findings on backward predictability, we therefore opted to focus on estimating forward predictability, which is consistent with the assumption of linearity in formulating speech (Levelt et al., 1999). Surprisal, defined as: Sðsyllable i Þ ¼ Àlog 2 Pðsyllable i jsyllable iÀ1 ; syllable iÀ2 Þ, was computed for the target syllables. The surprisal values were derived from a trained language model of the German web corpus (DeWaC) consisting of more than 1.34 Â 10 9 words from written texts in different genres (e.g., newspapers, chats). The choice of using a written corpus was motivated by the limitation of small-sized spoken corpora in representing rarer/infrequent syllable types (M€ obius, 2003;Schweitzer and M€ obius, 2004). Besides, there is evidence suggesting that written corpora provide comparable estimates of syllable frequency to spoken corpora for German (Samlowski et al., 2011). Syllable boundary assignment was fine-tuned, using the Hidden Markov Model (HMM) syllable tagger (Schmid et al., 2007). A syllable-based forward tri-gram language model was then generated, using the SRI Language Modeling Toolkit (SRILM) (Stolcke, 2002). According to the language model, target syllables were assigned with high and low surprisal values. A high surprisal syllable is less predictable (S % 5.05), whereas a low surprisal syllable is more predictable (S % 0.58). In total, we had 60 sentences, to be recorded in three noise conditions (reference ¼ no noise, 0 dB and À10 dB SNR).

C. Procedure
Participants were instructed to read aloud a set of stimuli at their habitual pace in a soundproof studio. They wore over-ear headphones (AKGK271 MKII, AKG Harman) and a head-mounted microphone (DPA 4067-F Omni, DPA Microphones), with a display computer in front. Sentence stimuli were orthographically presented in two lines in the centre of the computer screen. Note that some sentence stimuli were shortened to fit within the 2-line limit. Participants were informed about the presence of noise during reading. Noise was played through the headphones. Eleven practice sentences were provided for participants to become familiar with the procedure and for the research assistant to calibrate the equipment in the control room. The experiment consisted of three noise conditions in separate blocks, with the middle block reserved for the reference (no noise) condition. The order of the other two noise conditions (0 dB vs -10 dB) was counterbalanced across participants. This design was adopted in order to minimize the confounding linear correlation between speech modification and the presentation order of noise (from quiet to soft to loud) as observed in our pilot experiment using a subset of 30 DeWaC sentences. The experiment took about 30 min to complete. Productions were recorded and stored as a mono.wav file with a sampling rate of 48 kHz and 24 bits per sample.

D. Data annotation
A total of 1906 sentence stimuli were phonemically and orthographically transcribed, after removing 363 items for mispronounciation (N ¼ 152) or disfluency due to the presence of pause/hesitation (N ¼ 211). Words, syllables, and segments in each sentence were first automatically annotated using Web-MAUS (Schiel, 1999). Two trained phoneticians subsequently checked all automatic annotations, and manually adjusted the boundaries of the target CV syllables and their constituent segments using Praat (version 6.1.08).

E. Statistical analysis
Four syllable-level acoustic features (duration, average intensity, intensity range, and F0), and three vowel-level acoustic features (F1, F2, and F2-F1) were chosen as dependent variables, in accordance with previous research (Boril and Poll ak, 2005;Brandt, 2019;Brandt et al., 2019;Castellanos et al., 1996;Lu and Cooke, 2009). F0 and formants were measured at the mid-point of each vowel because visual inspection of the data revealed minimal trajectory differences across experimental conditions. Formant values were determined using the default setting of Formant(Burg) in Praat, adjusting the maximum formant value for gender (5 KHz for male and 5.5 KHz for female). All feature values were extracted using in-house Python and Praat scripts. Since the normality assumption was violated in a Shapiro-Wilk test (Shapiro and Wilk, 1965), syllable duration, intensity and F0 were transformed into z-scores per participant (Simpson, 2009;Traunm€ uller and Eriksson, 1994), and formants were also normalized per participant using the Lobanov method (Adank et al., 2004).
Each normalized acoustic feature was statistically analyzed by fitting linear mixed effects models (LME) using the lmer package (Bates et al., 2015) in R (R Core Team, 2018) and evaluating model fits using the lmerTest package (Kuznetsova et al., 2017). Fixed effects included surprisal group, noise conditions, and their interactions. Surprisal group was coded as a simple contrast, and noise conditions were Helmert contrast coded to create two comparisons: (1) absence vs presence of noise (i.e., noise condition 1), (2) 0 dB vs À10 dB SNR noise levels (i.e., noise condition 2). Covariates included target syllables, part of speech (PoS), and lexical frequency for word stimuli. Sentence stimuli were treated as a random effect. We first constructed a by-sentence random intercept model for each acoustic feature. This model was evaluated against models with intercept þ slope for fixed effects to identify the optimal random structure, on the basis of Akaike information criterion (AIC). Each covariate was then evaluated for inclusion. Significance of effects in each model was evaluated by performing maximum likelihood ttests using Satterthwaite approximations to degrees of freedom. Alpha was Bonferroni-adjusted for multiple pairwise comparisons.
These results suggest that formant modification is primarily attributable to the presence of noise and that formants are not modified to be peripheral. To check if the formant change is a side-effect of higher F0 or intensity as an adjustment to noise, we conducted a series of correlations. The results revealed significant positive correlations between average intensity and F1 for front (r ¼ 0.34**), central (r ¼ 0.13*), and back vowels (r ¼ 0.40***), but only a significant positive correlation between F0 and F1 for back vowels (r ¼ 0.10**).1 These patterns suggest that the overall F1 raising for the three vowel types might be due to increasing intensity.

IV. DISCUSSION
The present study has extended previous research by examining the combined effects of noise and local contextual predictability (i.e., surprisal) on the acoustic realisations of syllables in German. We hypothesized that syllables with high surprisal or in noise will be hyperarticulated when compared to their counterparts with low surprisal or without noise, and that this effect will be stronger when the noise level increases. In addition, we expected the magnitude of hyperarticulation to be larger for syllables with high surprisal than those with low surprisal in noisy conditions, if speakers choose to make the least predicted element intelligible.
Our results provided evidence for the expected effect of noise on hyperarticulation: longer syllable duration, higher average intensity, larger intensity range, and higher F0, higher F1, and smaller F2-F1 difference in the presence of noise than in its absence. In addition, noise levels increased average intensity, intensity range, and F0. Our results also revealed an effect of surprisal, resulting in longer syllable duration and larger intensity range for syllables with high surprisal. Contrary to our expectation, the effect of surprisal was not more pronounced in the presence of noise or at variable noise levels (as reflected in the lack of any interaction effects).
However, not all the measured acoustic features were manifested for the sake of increasing detectability or enhancing phonological contrasts. Although the syllablelevel features were acoustically enhanced to become more detectable in noise, vowel formants did not follow the same pattern. Previous studies have shown that vowel peripherality benefits intelligibility (Ferguson and Kewley-Port, 2007;Smiljanić and Bradlow, 2005). Therefore, we would expect lower F1 and higher F2 for closed front vowels, or higher F1 for open central vowels, or lower F1 and F2 for closed back vowels in a noisy condition.
Our results did not yield those expected patterns for the three vowel groups. On the contrary, the F1 of the front, central, and back vowels under investigation increased uniformly in a noisy condition. The F1 increase might be related to the corresponding increase in average F0, which could arise from tense vocal folds because previous studies have reported a moderate correlation between F0 and formants that stems from co-variation in the size of laryngeal and supra-laryngeal structures (Fant, 1970;Fitch and Giedd, 1999). Another possible explanation might be related to the observed higher average intensity and larger intensity range in noise, as previous studies have shown that the intensity contour co-varies with the mouth-opening area (Chandrasekaran et al., 2009;He et al., 2019). The higher F1 might then be a consequence of having a large openmouth area (i.e., analogous to jaw-lowering) to increase intensity in noise. To determine one of the postulated explanations, we examined correlations among F1, F0, and average intensity. The results yielded a positive correlation between F1 and average intensity, suggesting that F1 is an ancillary effect of increase in intensity.
Despite the consistent increase in F1 for front, central and back vowels, noise exerts differential effects on F2 of these vowels. While F2 lowers for front vowels (see Lu and Cooke, 2008) for similar observation), it does not statistically change for central or back vowels. The lower F2 for front vowels is not consistent with the idea that vowels are modified to be more peripheral in the vowel space for better perceptual distinction. This interpretation is further supported by the smaller F2-F1 difference across front, central and back vowels in a noisy condition. Note that F2-F1 difference is a derived measure by subtracting F2 from F1. Given the weak and selective effect of noise on F2, the overall F1 increase across all vowels could drive the resultant F2-F1 differences. It seems that noise does not enhance the peripherality of vowel distinctions.
The lack of vowel peripherallity in the presence of noise or with noise levels could also be due to the nature of our selected vowels. Three out of fvei vowels (/i:, a:, u:/) are point vowels. They are characterized by extreme F1 and F2 frequencies. These vowels are likely subjected to anatomical constraints within the oral tract as to the extent and direction that they can be made more peripheral. Such observation is in line with Wedel et al. (2018) showing that the manner vowels are hyperarticulated is vowel-specific in American English, rather than the result of an expanded vowel space per se. On top, German has a crowded vowel space, which might constrain possible acoustic modifications, in consideration of the need to preserve phonological vowel distinctions. From a perceptual perspective, it may be more appropriate to enhance the signal via intensity, duration, and F0 than increasing vowel peripherallity in response to noise.
As expected, noise levels (0 vs À10 dB SNR) significantly increased average intensity and F0 and expanded intensity range. Previous work investigating different noise levels shows similar results. For instance, Ngo et al. (2017) found an increase in intensity and F0 with the increase in noise levels. In contrast to Ngo et al. (2017), we did not find such effect on syllable duration or vowel formants. It seems that speakers are less likely to modify syllable duration or vowel-level features such as formants in response to noise levels. One potential explanation could be related to the temporal order of the manipulated noise levels in our experimental design. Speakers always recorded no-noise sentences in between the two noise conditions, which could serve as a "reset" of the expected incremental adaptation of noise levels as observed in Ngo et al. (2017).
Broadly, our findings of enhancing multiple acoustic features in response to noise suggest that speakers strategically choose to encode redundant acoustic signal to minimize errors in an unreliable channel or adverse acoustic conditions. Furthermore, the increase in noise levels leading to higher intensity and F0 suggests that speakers make a dynamic assessment of their environment. This finding is consistent with the "hyper"-and "hypo"-articulation (H&H) model of speech production (Lindblom, 1990), where speech is viewed as an adaptive system that is sensitive to the real-time contexts for speakers.

B. Effects of contextual predictability
Our study revealed that syllables with high surprisal not only exhibit longer duration but also larger intensity range than those with low surprisal. Conversely, our findings could also be interpreted as syllables with low surprisal to exhibit shorter duration and smaller intensity range. Irrespective of the direction of change, surprisal affects not only syllable duration but also intensity range, extending the findings of Aylett and Turk (2006). This is in general agreement with previous work on the effect of predictability for duration at the level of word (Buz and Jaeger, 2016), morpheme (Tang and Bennett, 2018), syllable (Aylett and Turk, 2006), phoneme (Bybee, 2002), or interactions between levels (e.g., Hasegawa-Johnson et al., 2009). Duration is shortened in more predictable contexts as evidenced in 319 different languages (Pimentel et al., 2021).
The observed syllable duration effect is in line with the idea that longer syllable duration reflects explicit encoding to improve intelligibility of hard to understand units see Jaeger, 2010 andGahl et al., 2012 at the word level. As previously mentioned, the syllable duration effect could come from shortening of predictable syllable or lengthening of unpredictable syllable. This seems to argue in favor of accounts that treat frequent syllables as a holistic phonetic motor plan/unit for ease of retrieval, relative to less frequent syllables which are computed on-line (B€ urki et al., 2015;Laganaro, 2019;Whiteside and Varley, 1998).
At the segmental level, previous studies have found that American English vowels were more centralized in contextually more predictable syllables (Aylett and Turk, 2006), vowel space was more dispersed when they were contextually less predictable (Malisz et al., 2018), and formant trajectory were generally affected by surprisal (Brandt et al., 2021). Contrary to the work of Aylett and Turk (2006);Brandt et al. (2021), andMalisz et al. (2018), surprisal does not alter the first two formants of the five vowels /i:, e:, a:, o:, u:/ in the current study. Our results would seem counter to previous findings, but such divergence could arise from the following notable differences between studies: namely, the way surprisal is estimated. For instance, Aylett and Turk measured syllable probabilities by taking "uni-gram," "bi-gram," and "tri-gram" into consideration to arrive at two new factors: wide context and narrow context redundancy. They then estimated "language redundancy" according to the distribution of these new factors. In Brandt et al. (2021) and Malisz et al. (2018), a phone-based LM was used to calculate surprisal. Counter to these studies, the current study estimated surprisal in terms of a syllable-based trigram model. Because of our choice to train the LM with syllable as a unit, such mixed findings raise further questions as to predictability being estimated through these different measures and how syllable-based predictability might interact with phone-based predictability. Moreover, the current study estimates forward predictability. When a listening condition is noisy, it is possible that a listener might not immediately commit to a linguistic unit, e.g., the syllable, in perception. If a speaker considers such a listener perspective in formulating speech, backward predictability might also affect the acoustic realisation of syllables, a topic that deserves further investigation.
There are also other possible explanations for the lack of surprisal effect on vowel formants. Our study is different from previous stated research in the nature of analyzed vowels. While Brandt et al. (2021) and Malisz et al. (2018) studied all the vowels in the vowel spaces of the investigated languages, our study focused mainly on point vowels which could have a dispersion ceiling effect as discussed previously. Another important constraint for vowel formants variability in our study is that all our syllables are stressed syllables, while Aylett and Turk (2006) analyzed both stressed and unstressed syllables. Stressed syllables will constrain the degree of vowel reduction on highly predictable syllables.

C. The combined effect of noise and contextual predictability
The present study was designed to investigate the combined effect of predictability and background noise on the acoustic realization of German syllables. Initially, we hypothesized that speakers will phonetically enhance high surprisal syllables more than low surprisal syllables in noisy conditions if speakers choose to make the least predicted element intelligible and detectable in noise.
In our study, only duration and intensity range could allow us to answer this question as they were subject to both effects of surprisal and noise (Fig. 1). Contrary to our expectation, this study did not find a significant interaction between surprisal and noise effects. Our results are consistent with Zhao and Jurafsky (2009), who investigated the effects of word frequency (another measure of predictability) and noise on the acoustic realization of Cantonese tones. Although they found a main effect of noise on all tone types and word frequency for Cantonese midrange tones, they did not find any interaction between noise and word frequency.
This parallel/additive effects of both noise and surprisal suggest separate, independent processing. The information theory framework (Shannon, 1948) could help us interpret those results. In Pate and Goldwater (2015), two types of signal encoding are distinguished: source code and channel code. While source coding focuses on message, channel coding is concerned with finding nearly optimal codes to transmit the message over a noisy channel with a low error rate (i.e., near the channel capacity). In our study, the predictability effect can be considered as message modulation (source coding) while the noise effect communication channel modulation (channel coding). The lack of an interaction finding suggests that the transmission channel is expanded (i.e., modified) without compromising source coding. Alternative explanations are also possible for the lack of interaction between surprisal and noise. One is related to the type of noise being experimentally manipulated. White noise might not have been adequate to degrade high-suprisal syllables to a larger extent than low-surprisal syllables. It is possible that noise manipulation using babble noise could have induced stronger degradation effects on the intelligibility of high-vs low-surprisal syllables to increase the chance of observing any interaction. The other explanation is related to the relatively small effect size of the interaction, compared to that of the main effect, suggesting that the test of interaction between noise and surprisal may be statistically underpowered in our study (see supplementary material). 1 Upon closer inspection of patterns from individual participants, only 8% of all participants (3 out of 38) showed significant interactions, suggesting that the interaction effect is not a preponderant group pattern. On this basis, we are more inclined to interpret our results as suggestive of the additive effects of predictability and noise. However, it is important for future studies to increase the sample size by collecting more data per speakers to further test the hypothesised interaction.
These results have implications for the need to go beyond message coding to include channel coding in formulating speech production models. Although channel coding is not part of linguistic representation (message formulation) during planning, it shapes the phonetic output. Our study has only explored one type of channel. Different types of channel abound, e.g., talking to L2 learners or robots, etc. Channel characteristics imply the need to consider contexts as part of an enriched formulation of phonetic output during planning.

V. CONCLUSIONS
Despite the variety of studies on speech enhancement strategies, these studies have focused on the Lombard effect or predictability effect separately while our work extends previous literature by examining both factors in tandem. The present study investigated whether surprisal, or contextual unpredictability, is modulated by different levels of background white noise. The presence of noise affected all syllable-based and most of the vowel-based metrics. Noise levels affected only syllable-based measures [intensity (average, range) and F0 but not duration]. Contextual predictability only influenced duration and intensity range. Contrary to our expectation, the effect of surprisal was not more pronounced in the presence of noise or for noisier levels, as reflected by the lack of any interaction effects. These findings suggest that speakers might aim at being maximally informative to avoid any potential mishearing, even when it entails articulatory effort to produce predictable syllables.