Nonmanual markers, such as head and eyebrow movements, eye blinks, and mouth shapes, are an important part of natural languages, both spoken and signed. Recent developments in computer vision have made it possible to extract facial and body landmark positions, as well as head-rotation measures, from 2D video recordings, which can be further processed to analyse the kinematics of nonmanual articulators. In this paper, we present an R-based workflow for processing raw outputs of computer vision toolkits with the goal of producing reliable and interpretable kinematic measurements of nonmanual articulators.

We compared four tools for analyzing blink velocity and amplitude, examining how MediaPipe, OpenFace, InsightFace, and 3DDFA compare in terms of blink analysis. Building on previous findings that different tools yield different results (Kuznetsova and Kimmelman, 2024), we explored their fixed-effect estimates across linguistic versus non-linguistic blinks, within non-linguistic blinks (eye watering blinks versus gaze-direction-change blinks), and within linguistic blinks (prosodic/turn-taking blinks, sign-aligned/list-marking blinks and backchanneling blinks), while controlling for head pose (Pitch, Roll, Yaw). Using mixed-effects linear models on annotated French Sign Language data, we found tool-specific patterns: consistent negative effects for InsightFace and MediaPipe, but positive. effects for 3DDFA. In addition, the influence of head pose varied across models (Pitch is strongly positive in MediaPipe but negative in InsightFace and some 3DDFA models; Roll and Yaw also switch importance across tools). These discrepancies highlight methodological biases that can distort linguistic interpretations.