Phonetics Encode Language Identity More Than Grammar

We test whether phonetic surface form alone can predict grammatical typology across languages. Using verse-aligned parallel Bible translations in 14 typologically diverse languages, we convert each verse to an International Phonetic Alphabet (IPA) character sequence (via Epitran) and predict two language-level WALS features: basic word order (SOV/SVO/VSO) and gender system size (None/Two/Three/Many). Because labels are constant within a language, naive random verse splits leak language identity and dramatically overestimate generalization; we therefore treat leave-one-language-out (LOLO) evaluation as the primary protocol. Across TF--IDF character n-gram baselines, phonological inventory/phonotactic features, BiLSTM and Transformer encoders, and a gradient-reversal adversarial objective that suppresses language-ID cues, random splits yield near-perfect accuracy, consistent with memorization. Under LOLO, performance is modest and highly variable across held-out languages, and embedding-space analyses show representations cluster by language identity (and correlated genealogy/area effects) more strongly than by typology labels. We release a reproducible IPA pipeline and provide an evaluation caution: in this phonetics-only setting, IPA robustly encodes language identity, while transferable signal for broad grammatical typology is weak and sensitive to label coverage.