Modeling cross-language vowel similarity: integrating phonetic analysis and explainable classification

Abstract

Abstract This study provides a multidimensional comparison of Standard Modern Greek and Standard Albanian, examines bidirectional cross-language vowel categorization, and identifies the acoustic cues that drive discriminability. Forty-two adults produced nonce words in carrier phrases; we extracted midpoint frequencies for the first three formants, formant dynamics, and vowel duration, and analyzed acoustic differences with mixed-effects models. We also simulated bidirectional category mapping using Random Forest classifiers trained on one language and applied to the other, and used explainable artificial intelligence (XAI) to identify the features driving classification. The results showed that the shared five-vowel inventory was comparable across languages in overall spatial overlap, though with selective spectral/dynamic differences and pervasive timing divergence. The mapping simulations yielded largely corresponding mappings for shared vowels in both directions, although the reciprocal mappings were not fully symmetrical. Albanian /ə/ was classified mainly as Greek /o/, and Albanian /y/ primarily as Greek /i/. The machine learning classifier performed well, with XAI indicating duration as the dominant cue. The study demonstrates that cross-language vowel similarity is inherently multi-dimensional, since apparent equivalence in canonical vowel space can coexist with systematic temporal and spectral divergence that supports robust discriminability and predictable category mapping.