Typing Rhythms Beyond the Latin Script: An Amharic–English Keystroke Dynamics Benchmark

Keystroke dynamics---the timing of a person's key presses and releases---is a low-cost, non-intrusive behavioral biometric. However, public benchmarks and most empirical evidence focus on typing in Latin scripts, predominantly English, and do not cover speakers of Ge'ez-script (Ethiopic) languages such as Amharic, which has more than 35 million native speakers. Ethiopic syllabic characters are commonly entered through multi-key transliteration, creating different motor-planning demands. This raises the question of whether timing-based verification methods developed for Latin scripts transfer effectively to other writing systems. We present and characterize a keystroke dynamics corpus collected from a shared pool of typists completing three text-entry tasks---fixed text, free text, and password entry---in Amharic and English. The corpus contains 114,955 labeled keystroke events across five sub-corpora (4,751--51,813 rows each), with up to 25 raw and derived timing fields per event. We report a systematic exploratory data analysis (EDA) covering structural composition; missingness and duplication; demographic and skill-level coverage; univariate and bivariate timing distributions; feature redundancy (correlation and hierarchical clustering); per-user sample imbalance; and user distinguishability in Principal Component Analysis (PCA), t-Distributed Stochastic Neighbor Embedding (t-SNE), and Uniform Manifold Approximation and Projection (UMAP) projections. The analysis finds strongly right-skewed timing features, a small number of algebraically redundant feature pairs, a coefficient of variation in per-user sample counts of up to 2.4, and substantial visual overlap between users in low-dimensional projections. Together, these properties plausibly explain the 40--60% rank-1 accuracy and 10--20% equal error rate (EER) previously obtained with baseline classifiers on this cohort. Based on these diagnostics and recent literature, we recommend feature engineering, session- and user-aware evaluation protocols, and careful model selection for keystroke biometrics in low-resource, non-Latin-script settings. We also make the EDA pipeline available for benchmarking. To the best of our knowledge, this is the first analysis of tri-modal keystroke dynamics data in an Ethiopic-script language, providing a reference point for cross-lingual and cross-script research in behavioral biometrics.