Camille Voudrin
Camille Voudrin (born 1988) is a French computational linguist and epigraphist, based at the Laboratoire d'Épigraphie Numérique et Linguistique Computationnelle (LENLiC) at the Institut national des humanités numériques (INHN) in Limoux, France. She studied mathematics and linguistics at the École Normale Supérieure before completing a doctorate at Paris-Diderot in 2015 on Hidden Markov Models applied to undeciphered writing systems. Her collaborators on the 2024 study, Solène Marchand and Hadrien Leclerc, are doctoral researchers at the same laboratory.
Voudrin's declared research programme is the development of "script-agnostic" methods: statistical tools that can characterise the internal structure of an unknown writing system without prior knowledge of the language, the sign inventory's size, or the script's typological classification. She has expressed particular interest in featural scripts—systems in which the graphic form of a sign encodes phonological properties—as a test case for whether structural methods alone can recover design logic.
Work
Voudrin, Marchand, and Leclerc published "A distributional test of vowel–consonant structure in an undeciphered signary suggests robust class separation" (Language Codes 7, March 2024: 1281–1294). The study applied a constrained two-state Hidden Markov Model to three extended inscriptions in the Kristiansen coding system, finding that approximately 90% of sign types fell at near-deterministic class probabilities (p(V) ≤ 0.05 or p(V) ≥ 0.95). The consonant-like class comprised box, corner, and triangle sign families; the vowel-like class comprised tee, line, meander, and barred-post families. Kristiansen's visual families showed zero cross-class mixing on a chi-square test.
The paper also derived a sonority hierarchy from onset and coda cluster directionality alone, consistent with obstruent > nasal > liquid > glide ordering, and identified C05 as the dominant cluster-initial element—a profile the authors described as "consistent with a coronal fricative or sibilant, though no phonetic claim is made."
Voudrin has been careful to frame the result as orthogonal to, rather than conflicting with, Rubergskier's arithmetic reading. "A text can simultaneously show word-level syllabic structure and clause-level arithmetic templating," she noted in a brief response to correspondence in Language Codes. "We are describing different levels of the same object."
The paper is notable for its methodological appendix, which provides full pseudocode for the constrained Viterbi algorithm and the sonority-learning procedure, and for nine diagnostic figures including vector tracings of the three corpus inscriptions used in the study.