There is a growing body of cross-linguistic research suggesting that co-speech gestures of the hands, arms, head, and other parts of the body are constrained in their timing according to various aspects of prosodic structure. In languages such as English, German, and Catalan, co-speech gestures are found to align with metrically prominent syllables, particularly stressed syllables that also bear phrase-level pitch accents (Loehr, 2012; Esteve-Gibert & Prieto, 2013; Kügler & Gregori 2023). In lexical tone languages, gestures also co-occur with syllables bearing metrical prominence, though prominence is often cued through very different means compared with stress-based languages. For example, in Medʉmba, a Grassfields Bantu language spoken in Cameroon, co-speech gestures are drawn to foot heads, which are cued through a broader range of consonantal and vocalic contrasts than non-heads (Franich & Keupdjio 2022; Franich et al., 2025). In Igbo, gestures tend to occur on word-final syllables (Franich, 2025), which have been analyzed as metrically prominent based on patterns of tonal downstep (Clark, 1990). Taken together, this research demonstrates that there is no universal acoustic or physiological cue (or set of cues) to which gestures consistently align across languages, making pure biomechanical accounts of co-speech gesture timing (e.g., Pouw et al., 2020; Pouw & Fuchs, 2022; see also Ambrazaitis & House, 2022) difficult to maintain.

The timing of co-speech gestures is also sensitive to the presence of prosodic phrase boundaries across multiple languages. For example, in Catalan (Esteve-Gibert & Prieto, 2013) and Medʉmba (Franich et al., 2025), findings indicate that gestures occur relatively earlier in a word when the word is followed by a phrase boundary. Esteve-Gibert and Prieto (2013) compare this behavior to that of pitch accents, which are also known to retract away from a phrase boundary (Silverman & Pierrehumbert, 1990; Prieto et al., 1995). Co-speech gestures are also found to lengthen in the vicinity of a phrase boundary, exhibiting a similar ‘clock-slowing’ tendency at the boundary to consonants and vowels (Krivokapić et al., 2014; 2017). Recent work has found that the timing of the onset of gesture strokes across languages correlates with the relative timing of cues to phrase-level prominence: in languages like English where focus-related durational lengthening influences both the consonant and vowel of the prosodically-prominent syllable (Cho 2006), gesture strokes start around the onset of the word. In languages where lengthening primarily influences the consonant and not the vowel, gesture strokes start before the word onset (Franich, 2025). These findings not only suggest a close relationship between prosodic structure and co-speech gesture timing, but also suggest that co-speech gesture timing is controlled as a part of the speaker’s prosodic grammar.

Prosodic phrasing is known to modulate word and syllable durations cross-linguistically (Turk & Shattuck-Hufnagel, 2007; Paschen et al., 2022; Kim et al. 2024), and we hypothesize that such effects may play a role in the timing of co-speech gestures in Igbo. However, variation in word and segment durations can also arise as a result of predictability effects, such as word frequency, lexical contextual probability (the probability of a word preceding or following another word), or phonotactic predictability (the predictability of a phone given a preceding or following phone) (Bell et al., 2003; Bybee, 2006; Ernestus, 2011; Pierrehumbert, 2001). While predictability has not been investigated explicitly for its relation to co-speech gesture timing, there are reasons to believe that predictability might be relevant for the planning and production of gestures. Predictability is known to influence word and segment durations: on the whole, more predictable words and segments are produced with shorter durations than less predictable words and segments (Section 1.3 for further discussion). Furthermore, and directly relevant to our Igbo data, there is the fact that both word frequency and lexical contextual probability increase the likelihood of coalescence patterns (Krug 1998).

Exactly how prosody and predictability might each influence co-speech gesture timing depends on the stage at which these phenomena influence the speech planning process. The locus of predictability effects within the speech chain is still a matter of some debate. Earlier work by Bybee & Hopper (2001) proposed that phonetic effects of predictability arise relatively late in the speech planning process as a result of the greater ease of retrieving articulatory plans for more frequently uttered words or sequences. In contrast, Aylett and Turk’s (2004) smooth signal redundancy hypothesis proposes that predictability effects arise at an earlier stage of planning. Specifically, they propose that speakers endeavor to achieve “robust communication,” i.e., the successful transmission of relevant phonological contrasts to the listener, while minimizing speaker and listener effort in contexts of language redundancy (where lexical, syntactic, semantic, and pragmatic factors conspire to make words more predictable in context). In this theory (see also Turk, 2010), phonetic effects of predictability are implemented in a prosodic planning component that optimizes information flow in the speech signal. Later work by Seyfarth (2014) locates phonetic effects of predictability within the speaker’s exemplar-based lexicon, based on the fact that the average lexical contextual predictability of a word influences its duration even when the same word occurs in more and less predictable contexts. In contrast with this account, Shaw and Tang (2021) argue that lexical representations are not directly influenced by predictability effects, but that language redundancy (in an Aylett and Turk-style model) influences prosody, which, in turn, can “leak” into the memories of words. This is based on their finding that lexical contextual predictability was associated with not only variations in duration, but also variations in fundamental frequency and intensity—two other correlates of prosodic prominence—in a corpus of Mandarin speech.

The place of predictability effects in speech planning has implications for the larger architecture of speech planning and production, including planning of co-speech gesture. If durational variations resulting from prosody and predictability arise from the same processing stage, there is greater likelihood that both will influence the timing of co-speech gestures. If, on the other hand, phonetic effects of prosody and predictability are controlled (even to some degree) at separate levels of planning, this leaves room for different effects of prosody and predictability on gesture timing. To take one recent model of speech production, Shaw and Tang (2023) propose that speech planning proceeds through a selection process in which three components—the lexicon (with phonetically-detailed exemplars), phonological component, and prosodic component—serve to influence the chosen production output simultaneously and in parallel. A different approach is taken by Keating & Shattuck-Hufnagel (2002), who see prosody as occurring at an earlier stage of speech planning relative to word form encoding. This model was designed to capture patterns such as the English ‘Rhythm Rule’ (i.e., changes in stress and/or accent location in the context of a clash, as in Japanése ápples → Jápanese ápples), which evidences a kind of prosodic restructuring of words that happens in response to variables such as prominence and phrasing. The authors propose that speech planning proceeds in stages, such that certain details about a word’s phonetic pattern are not necessarily present at earlier stages of prosodic planning:

Crucially for the present work, one way of viewing the variability of co-speech gesture timing in Igbo is through the lens of prosodic restructuring. As is discussed in Section 1, the planning of co-speech gestures is intricately connected with the process of prosodic planning. We propose that Igbo speakers, when anticipating prosodically-conditioned reduction processes, will ‘shift’ their co-speech gestures earlier in the word, just as speakers are observed to shift speech-based prominence in response to the prosodic context (see Section 1.4 for additional arguments for ‘gesture shift’ in Igbo). The model proposed by Keating and Shattuck-Hufnagel (in contrast to the model by Shaw and Tang, 2023) leaves open the possibility that the fine details of lexical representations—perhaps including those related to predictability effects—may not yet be available when prosodic restructuring takes place. As such, the model provides for the possibility that reduction effects linked to prosodic structure will have a different effect on the planning of co-speech gesture timing than reduction effects linked to predictability measures alone.

Aside from the influence of predictability on duration, another path by which co-speech gesture timing could be influenced by predictability effects is through the attraction of gestures to less predictable words and segments. Previous work has demonstrated for many languages that co-speech gestures are preferentially timed to words bearing new information in the discourse (Im & Baumann, 2020; Levy & McNeill, 1992; Gullberg, 1998; Muñoz-Coego et al., 2022; Franich, 2024). New information is also generally less predictable from the discourse context (Ferreira & Lowder, 2016). A further possibility is that gestures, instead of (or in addition to) being driven away from prosodically/phonetically weak syllables, are attracted to less predictable syllables. If this is the case, then predictability may be found to have an effect on the timing of co-speech gestures even if it does not influence word and segment durations in the expected ways. In Section 1.4, we provide further details on the factors that are known to condition gesture timing in Igbo, and outline how predictability may fit in as an additional explanatory variable in co-speech gesture timing. First, however, we provide further details on how predictability can influence speech timing.

While predictability has yet to be examined as a factor influencing co-speech gesture timing, there is a large body of existing work demonstrating its effects on speech timing. The term ‘predictability’ can encompass many different quantitative measures (see Shaw & Kawahara, 2018 for a recent overview), but generally characterizes the probability with which a given phone or word will arise in speech. While word frequency has long been linked to phonetic reduction (Bybee, 2006; Ernestus, 2011; Pierrehumbert, 2001; Zipf, 1929) and negatively correlated with word duration, specifically (Bell et al., 2009; Gahl, 2008), other forms of contextual predictability are demonstrated to have an effect on word and segment durations (Cohen Priva, 2015; Jurafsky et al., 2001; Seyfarth, 2014). The best studied of these effects is forward lexical predictability (FLP), which is the probability that a word will occur given the preceding word (Jurafsky et al., 2001; Krug, 1998; Saffran et al. 1996). FLP has been shown to be an even better predictor of word durations than raw lexical frequency, with words with greater FLP found to exhibit relatively shorter durations than those with lower FLP. Recently, research has suggested that backward lexical predictability (BLP)—the predictability of a word given the word that follows it—also exerts an influence on word durations similar to what is observed for FLP (Bell, 2009; Tang & Shaw, 2021). In addition to contextual lexical probability, contextual phonotactic probability has been explored in some languages. Evidence suggests that segments which are more predictable in their phonological contexts are more prone to lenition (Cohen Priva, 2015, 2017; Gahl et al., 2012; Shaw and Kawahara, 2017; Whang, 2018). And though most work to date has examined phonotactic probability effects on neighboring segments, various researchers have cited predictability as a functional motivation for phonological vowel harmony patterns, in which (typically nonadjacent) vowels within a word or a particular prosodic domain will agree in terms of their phonological features (Suomi, 1983; Kaun, 1995; Kimper, 2017). Indeed, McCollum (2020) finds that harmonizing vowels that are noninitial in the harmony domain in Kyrgyz are prone to phonetic reduction. Importantly, effects of phonotactic probability on reduction are demonstrated to depend on language-specific pressures around contrast preservation (Cohen Priva, 2017), meaning that not all languages are expected to show effects of the same measures of phonotactic predictability. For example, in a cross-linguistic study of lenition patterns, Cohen Priva finds that the specific segments that languages opt to lenite is dependent on how contextually predictable target segments are within each language.

Given these findings, one of the goals of the present work is to examine how various forms of predictability interact with speech and gesture timing in Igbo. As will be discussed, the timing of co-speech gestures is especially variable when words occur in phrase-medial position, an environment where words are particularly prone to reduction due, among other things, to a process of vowel coalescence in the language. Predictability measures have been found to increase the likelihood of coalescence, segment reduction, and deletion in previous work (Jurafsky et al., 2001; Krug, 1998; Saffran et al., 1996), leading us to believe that predictability effects may also have an influence on the timing of co-speech gestures. In addition to lexical predictability effects, we focus in on one specific variety of phonotactic predictability—vowel-to-vowel co-occurrence probability—which we hypothesize may be especially important for speakers of Igbo, a language with vowel harmony. We predict that vowel-to-vowel predictability will have an influence on word and segment durations, which, in turn, may influence the timing of co-speech gestures. We now provide more general details on gesture patterning in Igbo, with a particular focus on prosodic phrasing and tonal melody.

Igbo is a Niger-Congo language, part of the Volta-Niger subfamily (sometimes referred to as West Benue-Congo or Eastern Kwa) spoken in Nigeria by approximately 45 million speakers. One of the most extensive works to date on the phonology of Igbo is by Clark (1990). Among other things, Clark describes a process of floating tone-induced downstep in Igbo in the nominal associative construction which applies differently depending on whether the noun is disyllabic or trisyllabic: initial syllables of disyllabic words undergo downstep, whereas initial syllables of trisyllabic words resist it. Clark interprets this difference in behavior across the two types of nouns as related to their underlying metrical structures. She proposes that metrical prominence at the word level is assigned from right to left, with metrical prominence assigned to word-final syllables and alternating syllables moving leftward. The initial metrically prominent syllable in a trisyllabic word allows it to resist the tone docking that would trigger downstep in the associative construction. Metrical patterns for disyllabic high toned word ewu ‘goat’ and trisyllabic word akwụkwọ ‘book’ are given in (1). Note that Igbo also has a process of ATR vowel harmony, such that vowels within the word must generally match in ATR value ([–ATR] vowels are marked below with the underdot diacritic).

Independent evidence for the metrical status of word-final syllables in Igbo comes from our previous work on co-speech gesture timing, in which we show that word-final syllables are the most likely to co-occur with a gesture (Franich et al., 2025; see also Figure 1 below). However, word-final metrical prominence is not the only driving factor behind co-speech gesture timing in Igbo. We also find that gesture position is influenced by tone melody: specifically, words that contain a HH melody are more likely to have gesture aligned to the first high of the HH sequence. Thus, in a disyllabic word with a HH tone melody—attributable to a single, multiply linked high tone assigned from the left edge of the word (2)—gestures are significantly more likely to be attracted to the initial syllable of the word than the final syllable. Interestingly, in trisyllabic words with a HHH melody (again, attributable to a single multiply linked high tone), gestures gravitate to both the initial and final syllables, but less so to the medial syllable. This led Franich et al. (2025) to propose that the effects of tone melody are best explained through the mechanism of tonal feet, whereby binary groupings of high tones are assigned from left to right in the word; tonal feet can then compete with feet at the segmental level for the attraction of gestures. An example of a disyllabic HH word with both a segmental and a tonal foot is provided in (2); the position of the foot head at each level is underlined. For a trisyllabic HHH word, we propose that the word would be footed such that the initial two syllables form a tonal foot, and the third syllable initiates a new tonal foot, resulting in a parse of (HH)(H).

The relevance of tonal feet to gesture timing in Igbo is interesting in light of findings presented in Sections 3 and 4 on the effects of vowel co-occurrence probability on co-speech gesture timing. Briefly, our findings support the idea that co-speech gestures are avoided on phonologically redundant syllables; in other words, those that bear the same features as the preceding syllable. We discuss how tonal feet may fit into a broader pattern with respect to redundancy effects in shaping prosodic patterning.

Franich et al., (2025) also find an interaction between word and phrase position in predicting gesture timing, driven by the fact that words occurring in phrase-medial position are more likely to experience a “shift” in gesture timing from the word-final syllable to an earlier syllable in the word (initial position for disyllabic words, and medial position for trisyllabic words). Figure 1 shows results for the effect of phrase position on gesture timing in disyllabic and trisyllabic words in Igbo. Importantly, we find ample evidence that gesture timing varies within the same word according to phrase position. For example, one frequent word in the corpus, ạ̀nạ̀, meaning ‘land,’ consistently showed alignment of gestures to word-final position when it occurred in phrase-initial and phrase-final position, but showed alignment of gestures to word-initial position about half the time when it occurred in phrase-medial position. Given the observed variation within the same words, the effects of phrase position do not seem to be reducible to differences in the prosodic structure of the words under investigation.

Franich et al. (2025) hypothesize that the difference in patterning across phrase positions may be attributable to the susceptibility of word-final syllables to phonetic reduction processes, including (but perhaps not limited to) a gradient process of vowel coalescence that targets word-final vowels when they occur in hiatus before another vowel. Zsiga 1997 shows that this process—which is highly common due to the prevalence of words of shape VCV and VCVCV—can be modeled as a combination of final vowel shortening and articulatory overlap. She argues that the articulatory gestures (e.g., of the tongue and lips) of the word-final vowel are shortened in duration and heavily coarticulated with those of the following vowel.1

As discussed in Section 1.2, one interpretation of the effect of phrase position on gesture timing in Igbo is that words undergo metrical reorganization when word-final vowels are especially prone to reduction, and gesture shift reflects this reorganization. Reduction of a prominent syllable with an accompanying shift in prominence cues is not unexpected in the context of metrical patterning more generally: there are many documented cases of reduction-related stress shift across different languages (Lightner, 1972; Halle & Vergnaud, 1987; Al-Mozainy et al., 1985). Reduction of the word-final vowel in Igbo is also consistent with the typology of vowel coalescence and assimilation patterns across the world’s languages. Word-final vowels—which are the initial vowels in the V#V sequence undergoing coalescence in this case—should reduce more than the word-initial vowels that follow them (Casali, 1997). In summary, ‘gesture shift’ in Igbo bears many of the hallmarks of prosodic reorganization and stress shift cross-linguistically.

Given that coalescence in Igbo is a gradient phenomenon (Zsiga, 1997), and one that appears to be conditioned by factors that encourage vowel reduction, it is plausible that predictability could serve as an additional pressure leading to coalescence and ‘gesture shift’ away from final position in the word. If predictability effects are largely implemented through prosodic planning, as suggested by models such as Aylett and Turk (2004) and Tang and Shaw (2023), then we expect that they should have a similar effect on co-speech gesture timing as other prosodic variables that influence word and segment durations, such as prosodic phrasing. In other words, if, like prosodic phrasing, predictability leads to shortening of word durations and a greater incidence of vowel coalescence, then greater predictability may be related to a greater incidence of gesture shift to earlier positions in the word. We term this the duration reduction hypothesis. If, however, phonetic reduction effects related to prosodic phrasing and predictability arise from different stages in speech planning (as is possible in the model by Keating & Shattuck-Hufnagel, 2002), then co-speech gesture timing may be influenced by prosodic phrasing but not by predictability measures. We term this the prosodic reduction hypothesis. Finally, it is possible that predictability will exhibit effects on gesture timing that are less directly tied to duration shortening. For example, it could be the case that variation in gesture timing is partly driven by the attraction of gestures to less predictable, more information-rich syllables,2 which also happen to be less prone to durational reduction. In this case, vowel co-occurrence predictability may work in tandem with other prosodic factors in driving gestures away from word-final position when earlier syllables in the word are relatively less predictable/more informative. In this scenario, we would expect predictability measures, but not necessarily word durations, to have an influence on gesture timing. We call this the information hypothesis.

This research was granted ethics approval by the Institutional Review Board of Harvard University (Protocol IRB 22-1095). Four Igbo speakers from the greater Abuja area in Nigeria (3 men, 1 woman) participated in the study. The age range of the participants was between 35 and 55 years. Speakers were interviewed by a native Igbo speaker (the second author) in a quiet room or outdoor area. They were interviewed about aspects of language and cultural practices, such as traditions around marriage ceremonies and child naming.

Individual audio tracks were recorded for each participant using a Shure SM10 head-mounted microphone input to a Zoom Q8 video/audio recording device at an audio sampling rate of 48 kHz and a video sampling rate of 30 frames per second. Interviews were conducted for around 25–30 minutes, resulting in 9,070 words and around 25,000 phones recorded. Within this corpus, a total of 1073 gesture-aligned phones, or phones during which the manual apex of a gesture stroke occurred (see further details in Section 2.2), were annotated (annotation of gestures in the corpus is ongoing). Since the primary variable of interest was gesture position (by syllable) within the word, we exclude monosyllabic words from the present analysis. Furthermore, due to the relative sparsity of longer words in our corpus, as well as the tendency of longer words to be more complex in their structure, we limit our analysis to disyllabic and trisyllabic words. Our analysis focuses on the subset of these words whose tone melodies accounted for at least 2% of the total data, 478 words in all. The vast majority of the gestures obtained in the data (75%) were categorized as beat gestures, which constitute small, rhythmic hand and arm movements that typically align with prosodic emphasis but do not necessarily carry depictive meaning (e.g., as in the case of an iconic or a deictic gesture). We opt to include all gesture types in the present analysis in order to maximize our sample size, based on the fact that even iconic and deictic gestures are known to be constrained in their timing by prosodic structure (Shattuck-Hufnagel & Prieto, 2019; Rohrer et al., 2023).

Audio data were transcribed by a native speaker (the second author) and force-aligned using the FAVE aligner (Rosenfelder et al., 2022). Alignments were subsequently checked for accuracy by both authors. Manual gestures were coded by a team of trained annotators in ELAN software (ELAN 2024). Coding was carried out according to a set of criteria adapted from the MIT Speech Communications Group Gesture Studies Coding Manual (MIT Speech Communications Group, 2020). Coders kept the file’s audio muted to avoid auditory bias in coding decisions. Coders would observe which hand the speaker tended to gesture with most, and treated that hand as the dominant hand for the sake of coding. Coders were tasked with marking off intervals corresponding to communicatively intentional movements of the hands during communication. Movement intentionality is given by Shattuck-Hufnagel and Ren (2018) as a defining characteristic of co-speech gestures. Though the measure has yet to be defined kinematically, we note that it is likely indexed most closely by acceleration of the manual articulators (with faster acceleration for communicatively intentional gestures vs. fidgets). Part of the task of establishing reliability between coding partners in our study is accurately distinguishing “true” gestures from fidgets, based on impressions of intentionality. Our coders agreed over 95% of the time about which movements should be coded as gestures vs. fidgets in the dataset.

After identifying a gesture, coders marked various phases within it, including the preparation phase (if present), where the hands are brought into position before the gesture begins, the stroke, or the onset of intentional movement, any holds (pre- or poststroke, if present), where the hands rest in position after the stroke, and the recovery (if present), where the hands return to a rest position (Figure 2; Table 1). All these intervals were marked on a single ELAN tier. On a separate tier, coders marked off an interval within the stroke phase which corresponded with the gesture apex. In prior work, the definition of the apex has varied, with some authors using a spatial definition (usually the point of maximum extension of articulators such as the fingers in a hand gesture) (Loehr, 2012), while others have relied on a definition based on gesture timing, such as the peak or minimum velocity of movement of an articulator (Pouw & Dixon, 2019; Pouw, 2020). Given that peak velocity of the gesture has been found to closely align with prosodically prominent syllables, we opt to use peak velocity as our measure of apex timing. Apexes were coded in our data based on visual inspection, rather than through motion tracking data. Specifically, coders marked off the video frame in the data during which the hands moved the most, as reflected in visualized blurring of the hands across the video frame. We note that an alternative strategy of analyzing minimum velocity/peak extension of gestures did not yield substantially different results, likely because these landmarks tend to occur very close in time (at a latency much shorter than the average syllable). In prior work, we have demonstrated that this visualized point of peak velocity closely approximates the true moment of peak velocity as measured computationally (Dych et al., 2023).

Coders were also tasked with establishing reliability of the timing of their gesture apexes. To do so, pairs of coders were assigned the same participant’s file, annotating the file independently and then comparing their results. Their goal was to code such that the apex annotations in their respective versions of the file were within a single video frame. After establishing reliability (above 85%) on a series of video samples, coders in each pair worked on a participant’s file together, continuing to check that their judgments about gesture timing aligned, and discussing any disagreements that arose between them. Where consensus between coders could not be reached, a third research team member was brought in for a tie-breaking judgment.

Timestamps for gesture and speech data were extracted from ELAN .eaf files and Praat .TextGrid files, respectively, and merged using a Python script. Gestures were treated as ‘aligning’ with a particular phone if the apex of the gesture overlapped with any part of the phone. Each syllable of a gesture-aligned word was then coded for position in the word, position in the intonational phrase, the tone with which the syllable was realized, and the tone melody of the word as a whole. Cues to intonational phrases in Igbo are still not fully defined; as such, for the purposes of the present analysis, we use the edges of ‘breath groups’ (Lieberman, 1967) as an approximation of intonational phrase boundaries (Pierrehumbert, 1980). We observed a pattern of pitch reset—a common marker of an intonational phrase boundary across languages (Downing & Rialland, 2017; Kügler, 2017; Kula & Hamann, 2017)—at the beginnings of units consistent with most breath groups, giving us confidence that these breath groups were generally coextensive with intonational phrases.

Given prior evidence on the importance of lexical frequency and contextual lexical probabilities on word durations, we calculated both unigram frequency and forward and backward bigram probabilities for the words in our corpus. The equations for forward and backward bigram probabilities are given in (3).

Following prior work (Seyfarth, 2014; Tang & Shaw, 2021), lexical contextual predictability was calculated based on two language models which characterized the probability of each word in the corpus (containing a total of 9,070 words) given the word immediately preceding the target word or the word immediately following it. Language models were created using the kgrams package (Gherardi, 2024) for R statistical software (R Core Team 2024). Resulting model probabilities were then smoothed using the modified Kneser-Ney smoothing algorithm (Chen & Goodman, 1999) initialized with default parameters specified in the kgrams package. We note that the reliability of our language models will be somewhat limited compared with other studies given the small size of our corpus; unfortunately, given that tone in Igbo is rarely marked in orthographic representations, there is currently little existing data on which a more reliable language model could be built, since the functional load for tone is quite high in Igbo.

In addition to these measures of lexical predictability, we also computed a measure of phonotactic probability that we hypothesize should be important for Igbo speakers: vowel-to-vowel co-occurrence probability. Given Igbo is a language with vowel harmony, featural dependencies across vowels form an important part of speech production planning. Not only that, but speakers of languages that utilize vowel harmony rely on harmony patterns in word recognition in a way that speakers of languages without vowel harmony do not (Suomi et al., 1997; Vroomen et al., 1998). Furthermore, vowel co-occurrence predictability based on vowel harmony has been shown to lead to phonetic reduction (McCollum, 2020), a critical variable in our study. Here, we utilize two measures of vowel-to-vowel predictability: forward contextual predictability (the probability that a vowel will occur based on the preceding vowel) and backward contextual predictability (the probability that a vowel will occur based on the following vowel). For trisyllabic words, we focus specifically on predictability between the medial and final vowels within the word, since final vowels are those which are most targeted for reduction in Igbo.

In our previous work, we observed that co-speech gestures were more variable in their timing when they occurred in phrase-medial position. We hypothesized that this was due to the greater susceptibility of word-final vowels to coalescence and durational reduction in this phrasal position, but did not evaluate that possibility empirically. Here, we test that hypothesis by examining whether a) words are shorter when they occur in phrase-medial position compared with other positions; and b) word duration is a significant predictor of co-speech gesture position within the word. We also test the hypothesis that words will be shorter when predictability—including lexical frequency, contextual lexical probability, and vowel co-occurrence probability—is high, and that predictability will have an impact on gesture timing. As proposed in Section 1.4, the duration reduction hypothesis proposes that the causal chain between predictability, phrase position, and gesture shift is as follows: predictability and prosodic phrasing lead to shortening of the word, which makes it more susceptible to vowel coalescence. This leads to a shift in prominence, with an accompanying shift of a gesture away from word-final position. Under this scenario, both phrase position and predictability would condition vowel coalescence and gesture shift because they both contribute to reduction of word durations. As such, words that are both phrase-medial and more predictable will be most likely to undergo phonetic reduction, and will therefore be the most likely to exhibit gesture shift to syllables earlier than the word-final syllable. However, recall from Section 1.2 that certain models of prosodic planning propose that prosodic restructuring—the phenomenon hypothesized to be implicated in Igbo gesture shift—may not be sensitive to all types of phonetic variation. In the event that the phonetic effects of predictability are not planned by the time that restructuring takes place, we would expect effects of predictability on word durations, but not effects of predictability on gesture timing (consistent with the prosodic reduction hypothesis).

Duration reduction may not be the only path through which predictability could influence gesture timing. The information hypothesis states that gestures will generally gravitate to syllables that are less predictable and more informative, regardless of durational patterning. Recall from Section 1.3 that certain kinds of predictability measures, such as phonotactic predictability, do not affect segment durations in all languages. Nonetheless, we might still expect such measures to influence co-speech gesture timing. There are two ways that the information hypothesis could be supported in our data. First, it would be supported if gesture timing is sensitive to predictability measures but segment durations are not conditioned by these same measures. Second, it would be supported if gesture timing is sensitive to predictability measures, but the effect is demonstrably not driven by duration reduction. For example, gestures could be more likely to shift in words with high vowel co-occurrence predictability (a measure that is theoretically linked with vowel harmony patterns), even if these words are not phonetically reduced.

Related to this last point, we incorporate tone melody into our statistical models both as a way of controlling for known influences on co-speech gesture timing and to investigate possible interactions of tone melody and predictability in influencing gesture timing. In our previous work, we have shown that co-speech gestures tend to shift to initial position of a HH sequence (Franich et al., 2025). We have analyzed these effects as deriving from prominence asymmetries at the level of the tonal foot, in which the leftmost high-toned syllable is considered the head of the constituent. Nonetheless, HH sequences also embody an informational asymmetry: the second syllable in the sequence carries a feature which is redundant with that of the preceding syllable. To the extent that the information hypothesis can capture some of the variability in co-speech gesture timing in our data, we might expect that tone melody could interact with predictability measures in influencing gesture timing.

Three sets of statistical models were used to test our hypotheses. First, we constructed linear mixed effects models to investigate the effect of phrase position (three levels: initial, medial, and final) on both word duration and phone rate, with number of syllables (two vs. three) also included as a variable in the model. We opt to look at both word duration and phone rate in these models since they each provide a slightly different view of the effect of prosodic structure on speech timing. Word duration is a commonly used variable in the study of both prosodic and predictability-related effects on phonetic reduction (e.g., Baker & Bradlow 2009; Bell et al. 2003). However, prior work by Ryan (2016; 2019) has shown effects of prosodic ‘end-weight’—whereby heavier syllables and longer words tend to occur more frequently phrase-finally—which may be a confound in studying word duration effects. Phone rate is a measure of reduction that is robust to these effects, so it is useful as a complementary measure. An additional model investigated the interaction between phrase position, word position (again, with three levels: initial, medial, and final) and gesture presence (gesture vs. no gesture) on phone duration. For these three models, word and phone durations were log-transformed, and word and phrase position were sum-coded to allow us to compare how durations at each position compared with the grand mean. The next set of models examined the effects of the various predictability measures on word duration. All probability-related predictor variables were log-transformed, as was word duration. All continuous predictors were mean-centered to avoid collinearity (VIF values for fixed effects were all at 7 or below). Following previous work (e.g., Bell et al., 2009), we include two-way interactions between lexical frequency and our main predictability variables of interest, forward and backward lexical contextual probability and forward and backward vowel-to-vowel co-occurrence probability. We also include two-way interactions for our two lexical contextual probability measures and our two vowel-to-vowel probability measures. Finally, a mixed effects logistic regression model was used to investigate how phrase position, tone melody, predictability, and duration all contribute to the positioning of co-speech gestures within the word. The dependent variable here, word position, had two levels: final vs. nonfinal. For all but the final model (see Section 3.3), p-values were derived using Satterthwaite’s approximation. Post hoc comparisons were conducted in emmeans (Lenth et al. 2024) with p-values adjusted using Tukey’s HSD method. By-subject random intercepts were included for all models. Error bars on all plots reflect 95% confidence intervals.

Since no work to date has investigated effects of phrase position on word durations in Igbo, we first constructed a model investigating this effect. The model included phrase position and syllable count, as well as their interaction. Results are presented in Table 2. As expected, trisyllabic words had significantly longer durations than disyllabic words (p < .001). Differences in duration between words found in phrase-medial vs. phrase-final position were also found, with phrase-medial words realized with shorter durations, a significant difference from the grand mean (p < .05; Figure 3). Meanwhile, word-final syllables were realized with significantly longer durations compared with the grand mean (p < .001). Phrase-initial syllables did not differ significantly from the grand mean in their durations (p = .42). To ensure that the observed effects of phrase position on word durations was not simply an artifact of prosodic end weight, we also examined effects of phrase position on phone rate, measured as phones per second (Table 3). We find that phone rate for phrase-medial words is significantly higher than the grand mean (p < .01), whereas it is significantly lower than the grand mean for phrase-final position (p < .01) (Figure 4). Phone rate in phrase-initial position did not differ significantly from the grand mean (p = .91). Taken together, these results provide evidence for phrase-final lengthening in Igbo, consistent with findings from other languages (Edwards et al., 1991; Turk & Shattuck-Hufnagel, 2007; Berkovits, 1993; Katsika, 2016; Seo et al., 2019; Paschen et al., 2022; Kim et al., 2024). They also suggest, importantly, that phrase-medial position is subject to a greater level of phonetic reduction than phrase-initial or phrase-final position, consistent with the hypothesis proposed (but untested) in Franich et al. (2025).

Looking now at how duration interacts with gesture presence, we constructed a mixed effects logistic regression model aimed at evaluating whether word durations (log-transformed) predicted gesture position within the word (final vs. nonfinal). The model also included syllable count as well as an interaction between syllable count and word duration. Results indicated a significant effect of word duration on gesture position (p < .01; Table 4), as well as a significant effect of syllable count (p < .001). As seen in Figure 5, gestures were more likely to occur word-finally when word durations were longer. Though this relationship was observed to have a more positive slope for trisyllabic words, the interaction between word duration and syllable count did not reach significance (p = .10).

To better understand the relationship between durations and gesture presence, we also examined phone durations across different word and phrase positions with or without an accompanying co-speech gesture. In Figure 6, we see that while phrase-final syllables usually have longer durations than initial and medial syllables, this pattern changes in phase-medial position, where word-final syllables tend to have some of the shortest durations. This interaction between phrase position and word position was significant (p < .05; Table 5). These findings are again consistent with the idea that word-final syllables are reduced to a larger degree in phrase-medial position. It is also the case that syllables in all positions have longer durations in the presence of a gesture than when they do not carry a gesture (p < .001). This is consistent with the idea that syllables associated with co-speech gestures bear greater prosodic prominence than those that do not.3 Although collinearity effects precluded the incorporation of two- and three-way interactions in our model between word position, phrase position, and co-speech gesture presence, visual comparison of results in Figure 6 shows that the effects of co-speech gesture are not uniform across word and phrase positions when a gesture is present. Of note is the fact that word-medial syllables receive an especially large boost in duration in the presence of a co-speech gesture, precisely when the word occurs in phrase-medial position. This is consistent with the idea that gesture shift in phrase-medial position occurs in response to a shift in prosodic prominence within the word.

We now turn to look at various predictability effects on word duration within our corpus. A summary of these results is provided in Table 6. Four of our predictability measures—forward lexical probability (p < .05), forward V-to-V probability (p < .001), backward V-to-V probability (p < .05), and lexical frequency (p < .001)—were significant predictors of word duration in the model. As has been found in previous work, all three variables were negatively related to word duration. There was also a significant interaction between forward and backward V-to-V probability in predicting word duration (p < .001): specifically, word durations were even shorter where both forward and backward V-to-V probability were high. Finally, forward lexical probability and forward V-to-V probability both interacted with word frequency, but the effects were in opposite directions: word durations were shorter for words with both high forward V-to-V probability and high lexical frequency (p < .05). However, word durations were only shorter for words with high forward lexical probability when lexical frequency was low (p < .001). This latter finding may be attributable to differences between content and function words in terms of how predictability modulates word duration (most content words being less frequent, on the whole, than function words). For example, Tang and Bennett (2018) find that forward lexical bigram predictability was associated with shorter word durations for content words, but not for function words, in Kaqchikel Mayan. Such findings differ from prior work on English by Bell et al. (2009) who found a greater effect of backward lexical predictability for both content and function words. Tang and Bennett attribute their findings to differences in the morphological complexity of Kaqchikel words; we note that Igbo, like Kaqchikel, has a more agglutinating word structure than English, featuring a rich set of both inflectional and derivational affixes. Despite this typological variation, our findings in Igbo are generally consistent with cross-linguistic findings that greater predictability leads to shorter word durations.

The final model we present aims to examine how predictability effects influence co-speech gesture timing, alongside factors already known to influence gesture timing, including phrase position and tonal melody. Due to the large number of variables to be tested in the analysis and the potential for collinearity effects, we use forward nested model comparison to determine which factors contribute significantly to model fit. Starting with a model including only phrase position, tone melody, and word duration (number of syllables was omitted since this information is already coded in tone melody, e.g., HH vs. HHH), we added in each predictability measure to check for significant differences in model fit, measured through Chi-square tests of goodness of fit. Including lexical frequency did not lead to a significantly better fit of the model (χ²(478,1) = 1.854, p = .17), nor did inclusion of forward lexical probability (χ²(478,1) = 2.308, p = .13) or backward lexical probability (χ²(478,1) = 2.733, p = .10). Adding in forward V-to-V predictability did improve model fit (χ²(478,1) = 4.766, p < .05), as did adding in backward V-to-V predictability (χ²(478,1) = 12.045, p < .001). However, a model including both terms did not provide a significantly better fit than a model with only backward V-to-V predictability (χ²(478,1) = 0.404, p = .53); thus, forward V-to-V predictability was removed from the model. Model fit was significantly improved with the addition of an interaction between tone melody and backward V-to-V predictability (χ²(478,14) = 31.820, p < .01), but not between phrase position and backward V-to-V predictability (χ²(478,2) = 0.504, p = .78). In our last two model comparisons, we incorporated a) a two-way interaction between phrase position and word duration; and b) a two-way interaction between backward V-to-V probability and word duration in predicting gesture timing. Neither the first interaction (χ²(478,2) = 0.135, p = .93) nor the second interaction (χ²(478,1) = 0.740, p = .39) improved model fit.

Final model results are presented in Table 7. First off, as expected, both phrase position and tone melody are significant predictors of gesture timing (p < .001). Post hoc comparisons indicated that gestures were less likely to target word-final position when a word occurred phrase-medially vs. phrase-finally (β = –0.828, z = –3.025, p < .01) or phrase-initially (β = 1.261, z = 2.594, p < .05). As found by Franich et al. (2025), gestures are also less likely to target word-final position in words containing a single H tone span, as in a disyllabic HH or trisyllabic HHH melody. For example, gestures were significantly more likely to occur nonfinally in HH words compared with HL words (β = 1.436, z = 3.694, p < .05). Backward V-to-V probability was a significant predictor of gesture timing, with gestures less likely to target word-final vowels when this predictability measure was high. Interestingly, the significant interaction between backward V-to-V probability and tone melody indicated that gestures were even more likely to occur in nonfinal position when a word had both a HH sequence and had high backward V-to-V probability. Post hoc comparisons indicated a significantly less negative slope in the relationship between word-finality and backward V-to-V probability for HL words compared with HH words (β = 1.151, z = 3.113, p < .05), while differences in slopes were not significant between HL and LH words (β = 0.335, z = 0.946, p = .99) or between HL and LL words (β = 0.842, z = 1.054, p = .99). Though no significant differences emerged for slopes between different trisyllabic melodies, HHH words displayed numerically more negative slopes than HLH, HHL, and LHH melodies (the three other most common trisyllabic tone melodies in the dataset).

Figure 7 demonstrates the interaction of V-to-V probability and tone melody for disyllabic words. As can be seen, there was a stronger negative relationship between word position and backward V-to-V probability for HH words compared with words bearing other tone melodies. Figure 8, in contrast, presents the data with gesture position on the x axis and backward V-to-V probability on the y axis.

In sum, though our results in Section 3.1 support the idea of prosodic restructuring and gesture shift in phrase-medial position (where words are shorter and segments more reduced), results from Sections 3.2 and 3.3 do not provide support for the duration reduction hypothesis, as not all variables that influence word durations also impact gesture position. Notably, lexical frequency and forward lexical contextual predictability both had a negative influence on word durations; however, these factors did not play a role in influencing co-speech gesture timing. Furthermore, while backward vowel co-occurrence probability did influence both segment durations and gesture timing, our prediction of cumulative effects between predictability and phrase position was not borne out. More specifically, we anticipated an interaction between phrase position and V-to-V probability in predicting gesture timing, but this was not found, suggesting that these two factors influence gesture timing independently. Furthermore, taking a median split of the data for our backward V-to-V probability measure, it was not the case that words with higher V-to-V probability displayed nonword-final gestures more frequently in phrase-medial position. In fact, it was phrase-initial position that showed the greatest rate of nonword-final gestures among those words with the highest backward V-to-V probability. We now move on to discuss the implications of these results for models of co-speech gesture timing and speech planning.