Regardless of language modality, production rate can vary: words can be reduced in duration/length and the number of linguistic units produced over a time span can be higher or lower. But what conditions this variability? In this article, we look at lexical frequency as well as sociolinguistic variables, such as age, gender, region and language background in relation to sign duration, defined as the length in time of individual signs produced in context (i.e., not isolated production) and signing rate, defined as the number of signs produced per time unit within an utterance. We use corpus data from three unrelated sign languages, British Sign Language (BSL), Dutch Sign Language (NGT; Nederlandse Gebarentaal) and Swedish Sign Language (STS; svenskt teckenspråk) to research these issues cross-linguistically in the signed modality.

Zipf (1949) posited the law of abbreviation, which states that a word’s magnitude (i.e., its length) is negatively correlated with its token frequency: the more frequent a word is, the shorter it is expected to be. For example, frequent function words in English, such as the and it, are significantly shorter than low-frequency (e.g., domain-specific) words, such as agglutinating and grammaticalization (see, e.g., Bybee, 2007; Ernestus & Warner, 2011; Sigurd et al., 2004). This can be demonstrated with homophones, such as time and thyme (both pronounced /taɪm/), for which the more frequent time has significantly shorter duration than thyme when produced in speech (Gahl, 2008) – see also Wright (1979). The frequency–reduction correlation was initially attributed to the principle of least effort, such that language users will reduce the amount of articulatory effort to make communication more efficient (cf. Petrini et al., 2022). Work in usage-based linguistics attributes this type of reduction to routinization of production: phonetic strings that occur often are “practiced” more and thus quicker to produce (e.g., Bybee, 2007; Diessel, 2007; Lepic, 2019). Corpus-based cross-linguistic work on spoken languages has also attributed the frequency–reduction correlation to predictability and surprisal. The brain processes language based on contextual predictability, hence we are more likely to predict the next word of a sequence to be one that frequently follows the previous one(s), rather than one that rarely appears in that context (e.g., Gibson et al., 2019; Jurafsky et al., 2001; Mahowald et al., 2013; Piantadosi et al., 2011). In sign language communication, it has been shown that signs that are repeated in discourse are reduced in size and time when previously introduced, due to being established in the discourse and, thus, expected (Lepic, 2019; Stamp et al., 2024).

The length of a word is related to production rate, in that individual lexical items are produced faster or slower. The production rate (whether spoken or signed) may be influenced by many factors. Among spoken languages, there is plenty of research pointing to different sociolinguistic variables influencing production rate, e.g., age, gender and dialect (Byrd, 1994; Jacewicz et al., 2009; Keune et al., 2005). That speech rate is lowered as a function of age has been attributed to slower processing and control of the articulation apparatus (Horton et al., 2010; Jacewicz et al., 2009; Ramig, 1983). Speech rate is also part of the perceived age of a speaker, such that we identify lower rates as a property of aging (Skoog Waller et al., 2015). In American English, both gender and dialect have been found to correlate with speech rate, with men speaking at a higher rate than women, and certain regions having higher speech rates than others (Byrd, 1994).

To date, few studies have looked at production rate in sign languages. It has been argued that signing rate is lower than speech rate, measured as signs/words per minute, and that sign duration is reduced as signing rate increases (Grosjean, 1979; Wilbur, 2009). Higher signing rate can also lead to phonetic reduction, such as lowering the place of articulation of signs with higher locations in their citation form (Tyrone & Mauk, 2010). In a corpus study on STS, Börstell et al. (2016) found that sign duration decreases as a function of the sign’s token frequency, conforming to the law of abbreviation. Additionally, Börstell et al. (2016) found an effect of age on sign durations, such that durations increase with signers’ age, but found no effect of gender.

It is well known that there is substantial sociolinguistic variation in the lexicons of sign languages, conditioned by factors like age, gender and region (Bickford, 1991; Kimmelman et al., 2022; Lucas et al., 2009; Mudd et al., 2020; Safar, 2021), including the languages of this study (Börstell & Östling, 2016; Schembri et al., 2018; Stamp et al., 2014; Vermeerbergen et al., 2013). However, little is known about the potential effect of such sociolinguistic variables on duration and signing rate, other than what was found by Börstell et al. (2016) in terms of age (positive correlation) and gender (no effect). In many signing communities, there are perceptions about “who signs in what way” (e.g., Rowley & Cormier, 2023) – for example, some signers of BSL anecdotally report that Glasgow signers sign faster than others.

Here, we explore duration and signing rate on the basis of the following research questions:

We expect token frequency to be negatively correlated with sign durations, based on the general law of abbreviation (Zipf, 1949) and previous work by Börstell et al. (2016) on STS. We expect age to be positively correlated with duration and negatively correlated with signing rate, on the assumption that older signers sign more slowly, following previous work by Börstell et al. (2016) on STS, as well as work on spoken languages (e.g., Horton et al., 2010). For the other variables, we do not have any predictions of possible effects nor their directionality, other than the possibility of Glasgow signers signing faster among BSL signers.

For this study, we use corpus data from three sign languages: BSL (Schembri et al., 2017); NGT (Crasborn et al., 2015); and STS (Mesch et al., 2012, 2014; Öqvist et al., 2020). Sign language corpora are comparatively small, but the three corpora used here constitute some of the largest sign language corpora currently available (cf. Börstell, 2022b; Kopf et al., 2022). The data consists of ELAN annotation files (Crasborn & Sloetjes, 2008; Wittenburg et al., 2006) with segmentations for each sign produced in video-recorded sessions.

The data processing consisted of several steps. First, the ELAN (.eaf) annotation files and metadata files from the three sign language corpora were downloaded and read using R v4.4.1 (R Core Team, 2024) and the packages signglossR v2.2.6 (Börstell, 2022a) and xml2 v1.3.6 (Wickham, Hester, & Ooms, 2023). Only narrative and conversational text types were included, in order to focus on naturalistic and cohesive signing, thus excluding interviews and lexical elicitation tasks. The data was further processed, analyzed and visualized using the packages here v1.0.1 (Müller, 2020), tidyverse v2.0.0 (Wickham et al., 2019), afex v1.3.1 (Singmann et al., 2024), marginaleffects v0.21.0 (Arel-Bundock et al., Forthcoming), patchwork v1.2.0 (Pedersen, 2024), scales v1.3.0 (Wickham, Pedersen, & Seidel, 2023) and sf v1.0.16 (Pebesma & Bivand, 2023). We use the gloss annotations from the data, which are the individually segmented and labeled signs (i.e., lexical items). For the BSL and NGT corpus data, double annotations for two-handed signs were removed, such that each sign is only represented once, whether one- or two-handed. Empty annotations without a gloss were excluded, and signs with a duration of 0 were adjusted to a minimum of 1 millisecond. The three corpora use different criteria for determining the start of signs: the BSL Corpus defines start as the beginning of a transitional movement away from the previous sign, whereas the NGT and STS corpora define it as the start of the articulation phase.1 Thus, the durations and signing rates reported here can be compared within languages, but not necessarily across languages. After these steps, the resulting dataset contains approximately 50 hours of annotated data (BSL: 12.4 hours; NGT: 14.7 hours; STS: 23.2 hours), comprising 344,869 sign tokens (BSL: 59,925; NGT: 95,897; STS: 189,047) across 288 signers (BSL: 171; NGT: 75; STS: 42) representing different genders and ages (Figures 1, 2) and geographic regions (Figure 2). The sign tokens were further grouped into utterances. Utterances were split at any pause between signs of ≥500 milliseconds, a threshold selected on the basis of the approximate range of pause durations reported for spoken languages (see Heldner & Edlund, 2010). Utterances were used to calculate signing rate, defined as the number of signs divided by the total duration of the utterance (scaled to signs per minute). Only utterances with a length of ≥2 signs were included, so as to differentiate the metric from individual sign durations. In total, 21,358 utterances were segmented and included (BSL: 3,047; NGT: 6,707; STS: 11,604).

The data was combined with available metadata: for BSL, metadata about the signer and task is embedded in the file names; for NGT and STS, metadata is available in the repository of the respective corpus collection. Unlike BSL and NGT, signer age for STS is given as age group in approximately 10-year increments, and age for STS was thus estimated as the midpoint between the lower and upper bounds of each age group range. Age was treated as a continuous variable in the statistical modeling, but with less precision for STS. For BSL and STS, the metadata contains information about the family of the signer; NGT metadata contains information about the age of exposure to NGT. The variable family thus refers to whether or not the signer comes from a deaf or hearing family for BSL and STS (deaf family: BSL: n = 67 [39%]; STS: n = 15 [36%]). For NGT, age of exposure refers to the reported age of first exposure to sign language for each signer, defined as a continuous variable in years (M = 2.79; SD = 2.83).

In order to evaluate the effects of individual variables, a mixed effects model was constructed for each language, with sign duration (log-scaled and z-scored) as the outcome and sign token frequency (log-scaled and z-scored), age (centered by language), gender, family2 and region as fixed effects, with interaction between age and gender, and signer and sign as random effects. The fixed effects of these models were evaluated using likelihood ratio tests against a null model for each effect, showing that token frequency and age were the only significant effects across languages, and the only other variable that had a significant effect was region, for BSL only (Table 1). Based on these evaluations, a final model was constructed for all three languages together, with sign duration (log-scaled and z-scored) as the outcome and sign token frequency (log-scaled and z-scored) and age (centered by language) as fixed effects, with language, signer and sign as random effects. With this model, the effect of frequency was significant and negative (β = –.295; t(13530) = –40.095; p < .001***) and age was significant and positive (β = .006; t(253.1) = 8.778; p < .001***) – that is, duration decreases with sign frequency, but increases with signer age.3

Turning to signing rate, a mixed effects model was constructed for each language, with signing rate (log-scaled and z-scored) as the outcome and utterance length (log-scaled), age (centered by language), gender, family (BSL and STS only), age of exposure (NGT only) and region as fixed effects, with interaction between age and gender, and signer as a random effect. Table 2 shows the statistics of the individual fixed effects as compared against a null model, using likelihood ratio tests for each effect. Age again shows a significant effect across languages, as does utterance length. As with the duration models, only the BSL model shows a significant effect of region. A final model was constructed for all three languages, with signing rate (log-scaled and z-scored) as the outcome and utterance length (log-scaled) and age (centered by language) as fixed effects, with language and signer as random effects. With this model, the effects of utterance length (β = –.06; t(21350) = –10.119; p <.001***) and age (β = –.01; t(227.1) = –7.028; p < .001***) were significant and negative – that is, signing rates decrease as utterance length and signer age increase.

Figure 3 shows the observed duration and signing rates across languages, by age group (for visualization purposes; models used continuous age).

Across the three languages, only BSL showed an effect of region on both sign duration and signing rate. Figure 4 shows the predictions of duration and signing rate based on the individual BSL models, predicted for four ages from 20 to 80 years old. Figure 4 shows that the Birmingham signers are among the top for both models, indicating shorter durations and higher signing rates than the others (i.e., faster signing). At the other end, Cardiff signers end up with the longest durations and second to lowest signing rates, despite Cardiff having younger than average signers in the data (cf. Figure 1).

In this study, we looked at aspects of sign duration and signing rate across three sign languages on the basis of corpus data.

First, we were able to corroborate previous work, in that a sign’s token frequency is negatively correlated with its duration across all three languages. Although this finding is unsurprising, it shows how the law of abbreviation is found in the signed modality, as it has been found in spoken and written modalities.

Second, we showed that signer age correlates with sign duration (positively) and signing rate (negatively) across these languages, such that an increase in age results in longer sign duration and lower signing rate. This aligns with findings from spoken languages (e.g., Horton et al., 2010; Jacewicz et al., 2009; Ramig, 1983), and can be attributed to changes in both processing and physical production (i.e., articulation). An additional aspect to investigate in the future would be whether signing rate changes based on the age of the addressee and/or the difference in age between interlocutors in a dyad, as has been shown for spoken language (see Cohen Priva et al., 2017). That is, do younger signers accommodate their signing rate to older signers, or vice versa (cf. Stamp et al., 2016)? This question could not be addressed in this study, due to limitations in the size and stratification of the data, as it would require a more balanced sample of signers in different age group dyad combinations. In an ideal dataset for this research question, signers would be paired with younger, older and same age interlocutors in different conditions. This is not the case in our data, where most signers are paired with exactly one other signer. Furthermore, a more detailed investigation of pausing and the length and structure of sentences, utterances and turns may potentially show additional age-related differences across signers.

Third, whereas sociolinguistic variables, such as dialect and gender, have been found to influence speech rate in spoken languages (Jacewicz et al., 2009), and are known to affect lexical variation in sign languages (Bickford, 1991; Lucas et al., 2009; Schembri & Lucas, 2015), there was no significant effect of gender on duration or signing rate for any language in our study, which corroborates previous work by Börstell et al. (2016) on sign duration in STS. We additionally looked at signers’ language background with regard to family and age of exposure to a sign language. While “native signer” status is sometimes claimed to influence sign language production/perception in linguistic studies, the issue of “nativeness” is particularly problematic for sign languages, where the majority of signers tend to come from hearing, non-signing families, and, thus, are not strictly “native signers” in the traditional definition (cf. Cheng et al., 2021; Quer & Steinbach, 2019). In this study, we can conclude that the datasets reflect the general trend in the corresponding deaf communities of having a majority of signers from hearing families, but that neither family nor age of exposure seems to influence sign duration or signing rate.

Lastly, region had an effect on sign duration and signing rate only for BSL. Here, it was found that Glasgow signers have shorter sign durations than average, as expected on the basis of anecdotal perceptions in the BSL community, but this prediction was not corroborated by the signing rate model. However, seeing as the signing rate dataset is much smaller than the sign duration dataset, there are potential issues with the stratification of data across all ages and regions. For example, the number of signs and utterances for each of the regions Belfast, Glasgow and Cardiff is an order of magnitude smaller than that of Birmingham, Bristol, London and Manchester. This means the data and model are less reliable for the regions with fewer data points, which is also reflected in the range of the confidence intervals in Figure 4. Thus, while region may be a significant factor for sign duration and signing rate in BSL, we are hesitant to make any definitive claims about individual regions based on these data and analyses alone. Additional data may facilitate a reanalysis of possible differences between duration and signing rates across regions. Another approach is to address perceived signing rate by experimentally varying duration of individual signs and overall signing rate in different combinations, and have participants view and provide ratings for different conditions. We leave these possibilities for future work.

In summary, apart from seeing frequency effects in sign durations, we also see an effect of aging on both sign duration and signing rate across three unrelated sign languages. Aging as an effect on production rate is well established in spoken language research (e.g., Horton et al., 2010; Jacewicz et al., 2009; Ramig, 1983), and can now be shown also across different sign languages. In the bigger picture, the results of this study shed new light on the question of production rate as a phenomenon across modalities. The spoken and signed modalities come with different affordances in terms of the transmission channel and the main articulators employed. This may, in turn, lead to differences, such as the possibility for more simultaneous expression and a higher degree of iconicity in the signed modality. In principle, this could – as one reviewer points out – lead to restrictions on the amount of reduction that is possible, but yet we find the same effects as for spoken languages, in that frequency leads to articulatory reduction (shorter duration). Thus, our findings point to important similarities in how language is shaped across modalities, regardless of the way in which it is expressed. This is seen in how words/signs of a language compare to each other with regard to frequency of use, with reduction as part of efficiency, expectedness and routinization in communication, as well as in how the physical aging of speakers/signers alters the way language is physically produced and processed.