Empirical Investigation of Users' Preferred Timing Parameters for American Sign Language Animations
Sedeeq Al-khazraji, Becca Dingman, Matt Huenerfauth · 2020 · Extended Abstracts of the 2020 CHI Conference on Human Factors in Computing Systems (CHI EA '20) · doi:10.1145/3334480.3382989
Summary
This CHI 2020 Late-Breaking Work (7 pages) investigates a narrow but consequential HCI question: what timing values should be used when generating American Sign Language (ASL) animations from a script, so that Deaf viewers find them comfortable to watch? ASL is a primary language for roughly 500,000 people in the United States, and average English reading literacy among Deaf adults is lower than among hearing peers, so adding ASL to web content would meaningfully improve information accessibility. Recording human signers for every page update is impractical, which motivates ASL animation from virtual signers. Prior work by the same group had built data-driven timing models from motion-capture recordings of human signers and noted that users generally prefer ASL animations to be slightly slower than human video; but no prior study had systematically varied each of the five key timing parameters individually to identify preferred values. The five parameters are: sign duration (how long each sign lasts), transition time (hands moving between signs), pausing frequency (how often pauses occur, as a percentage of inter-sign locations), pausing length (seconds per pause), and differential signing rate (the exponent that magnifies or dampens natural speed variation across a passage, e.g. slowing at sentence ends). Sixteen native ASL signers (8 women, 8 men, median age 22.5, all learned ASL in childhood, 13 Deaf / 1 HoH / 2 CODA) watched 75-word passages rendered with Sign Smith Studio as five side-by-side variants per parameter and gave 1-5 preference scores. Statistical testing used Kruskal-Wallis followed by pairwise Wilcoxon with Bonferroni.
Key findings
Users preferred a sign duration of 1.62 seconds (SSS default), slightly slower than the 1.28 seconds observed in human signing. Transition time was preferred between 0.25 and 0.5 seconds, close to the 0.23 seconds typical of human signers. Pause durations of 0.125, 0.25, or 0.5 seconds all scored similarly and were preferred, compared with the 229ms average in human data. Pausing frequency peaked at 14% of inter-sign locations — notably less than the 25% pause rate observed in human signers, one of the more surprising findings. Differential signing rate was preferred at exponent 1 — meaning participants wanted the natural human-derived tempo variation applied directly, not dampened or exaggerated. Across all five parameters, no single value deviated dramatically from human signing, but the pattern is that animated signers should pause slightly less often and use slightly longer sign durations than their human counterparts. The authors attribute this to Deaf viewers needing a small amount of extra processing time when watching a virtual signer whose fine-grained facial and body movements are less expressive than a human.
Relevance
For accessibility practitioners building or evaluating signing-avatar content, this paper gives concrete numerical defaults to plug into a generation pipeline: 1.62 s sign duration, 0.25 s transition, 14% pausing frequency, 0.125-0.5 s pause length, and a linear (exponent 1) differential rate model. Those defaults are especially useful for organisations considering adding ASL to websites without recording human signers for every content update — a workflow that could make ASL on the web tractable for high-volume sites (news, government, e-commerce) where video re-recording is prohibitive. Limitations are substantial: sample size of 16, late-breaking-work format, each parameter varied in isolation (so interaction effects between parameters are unknown), no comparison against the human-equivalent value as a stimulus level for any parameter (making direct human-vs-animation claims indirect), and reliance on Sign Smith Studio, an animation authoring tool that has since been discontinued. The paper is useful alongside the Huenerfauth group’s fuller ASSETS publications on ASL timing models.
Tags: american sign language · sign language animation · signing avatar · deaf and hard of hearing · web accessibility · natural language processing · animation